DTM Generation Using Airborne LIDAR Data

sensorsReviewState-of-the-Art: DTM Generation Using AirborneLIDAR DataZiyue Chen 1,*, Bingbo Gao 2 and Bernard Devereux 31 College of Global Change and Earth System Science, Beijing Normal…

sensorsReviewState-of-the-Art: DTM Generation Using AirborneLIDAR DataZiyue Chen 1,*, Bingbo Gao 2 and Bernard Devereux 31 College of Global Change and Earth System Science, Beijing Normal University, 19 Xinjiekouwai Street,Beijing 100875, China2 Beijing Research Center for Information Technology in Agriculture, Beijing Academy of Agriculture andForestry Sciences, Beijing 100097, China; [email protected] Department of Geography, University of Cambridge UK, CB2 3EN Cambridge, UK; [email protected]* Correspondence: [email protected]; Tel.: +86-10-588-018-22Academic Editor: Vittorio M. N. PassaroReceived: 25 October 2016; Accepted: 24 December 2016; Published: 14 January 2017Abstract: Digital terrain model (DTM) generation is the fundamental application of airborne Lidardata. In past decades, a large body of studies has been conducted to present and experiment a varietyof DTM generation methods. Although great progress has been made, DTM generation, especiallyDTM generation in specific terrain situations, remains challenging. This research introduces thegeneral principles of DTM generation and reviews diverse mainstream DTM generation methods.In accordance with the filtering strategy, these methods are classified into six categories: surface-basedadjustment; morphology-based filtering, triangulated irregular network (TIN)-based refinement,segmentation and classification, statistical analysis and multi-scale comparison. Typical methodsfor each category are briefly introduced and the merits and limitations of each category arediscussed accordingly. Despite different categories of filtering strategies, these DTM generationmethods present similar difficulties when implemented in sharply changing terrain, areas with densenon-ground features and complicated landscapes. This paper suggests that the fusion of multi-sourcesand integration of different methods can be effective ways for improving the performance ofDTM generation.Keywords: DTM generation; surface-based; morphology-based; TIN-based; segmentation andclassification; statistical analysis; multi-scale comparison1. IntroductionIn past decades, the processing and applications of airborne Lidar (Light detection and ranging)data have been increasingly studied. Due to its high resolution in both horizontal and verticaldirections, airborne Lidar data can be employed for monitoring the change of landscape configurations,establishing building structures, analyzing tree volumes and creating 3D urban models. Althoughapplications of Lidar data vary, these subjects are built around one necessary procedure: the generationof digital terrain models (DTMs) using raw Lidar point clouds.Raw Lidar point clouds include ground and non-ground points. Through interpolation, the entirepoint cloud can be transformed to a digital surface model (DSM) whilst the ground points can betransformed into a DTM (Although “DTM” is a frequently used term in specific papers, the term“DEM” (digital elevation model) is sometimes employed by researchers to define the surface createdusing ground points. This terminology strategy may be a potential risk to a broad readership. DTMrefers to the bare earth surface, whereas DSM refers to a model that corresponds to the elevation of thesurface of man-made or natural objects (such as building and trees) and, if no such objects exist, thebare earth. DEM is thus a more generic term that could represent DTM, DSM, or any other elevationSensors 2017, 17, 150; doi:10.3390/s17010150 www.mdpi.com/journal/sensorsSensors 2017, 17, 150 2 of 24models. With the growing use of the term “DSM” in recent Lidar studies, it is recommended to employ“DTM” for specifically describing the terrain surface generated from raw point clouds). As a key stepof Lidar data processing, the quality of DTMs generated from raw point clouds not only influences theaccuracy and visual effects of these models per se, but also decides the reliability of other productsbased on these DTMs, such as nDSM (normalized digital surface model, DSM-DTM), individual treeand building models and land cover maps. Therefore, it is of both theoretical and practical significanceto propose effective algorithms for DTM generation. In the past two decades, many DTM generationalgorithms have been developed. These methods aim to produce DTMs from different perspectives,such as block-minimum, slope operator, triangulated irregular network (TIN) modelling and rastercalculation. Some DTM generation methods have been intendedly designed for such specific landscapetypes as forests or urban areas, whilst other algorithms are proposed for DTM generation in generallandscape types.To examine the performance of DTM generation methods under different circumstances, Sitholeand Vosselman [1] carried out a comparative experiment to test eight classic DTM generation methodsin 15 sampling sites. This study works as important reference for choosing appropriate DTM generationapproaches according to specific terrain situations. In addition, this project has collected standardreference data and a variety of sample data, based on which researchers can experiment, evaluateand compare their own algorithms with existing methods. The ISPRS (International Society forPhotogrammetry and Remote Sensing) sample data (http://www.itc.nl/isprswgIII-3/filtertest/index.html) has become one of the most important sources for experiments and accuracy assessmentsince 2003 onwards.Although a large body of algorithms has been proposed, DTM generation is still challenging [2–4].DTM generation methods are usually applied to large-scale sites. Therefore, it is very difficult to use aset of limited parameters for separating a complexity of terrain relief from a variety of non-groundfeatures. This explains the reason why DTM generation in urban areas is particularly difficult. Sitholeand Vosselman [1]’s experiment examined the suitability of mainstream DTM generation methodsand promoted the research of DTM generation significantly. Inspired by previous studies, many newDTM generation methods have been proposed and examined using the ISPRS sample data in the pastdecade. Additionally, some scholars [5–7] further concluded recent development of Lidar-based DTMgeneration. Zhang and Men [5] concluded advantages and disadvantages of some mainstream DTMgeneration algorithms, and proposed five recommendations for better DTM generation: adaptivethresholds, combination of different methods, multiple sources, complex terrain filtering and advancedclassification methods. However, this research did not provide detailed explanations and feasibleapproaches to implement the recommendations. Liu [6] reviewed recent development of Lidar systemsand processing algorithms, and discussed some critical issues: Lidar data filtering, model selection,interpolation methods, DTM resolution and Lidar data reduction. Although this paper introducedseveral categories of DTM generation algorithms, only a small proportion of DTM generation methodswere proposed since 2003. Meng et al. [7] conducted a comprehensive review of existing groundfiltering algorithms. This is by far the most recent and detailed review of Lidar-based DTM generation.Slightly different from previous nomenclature, this research classified six groups of ground filters:segmentation and cluster-based filters, morphological filters, directional scanning filters, contour-basedfilters, TIN-based filters and interpolation-based filters. This research illustrated typical algorithmsfor each category, and analyzed the advantages and disadvantages for specific filters. Meng et al. [7]further examined the performance of ten mainstream ground filtering algorithms in three types ofterrain: sites with rough slope and dense vegetation, sites with relatively flat urban areas containingobjects of various sizes and shapes, and sites with rough terrain and discontinuous surfaces. Based onthe comparison results, Meng et al. [7] suggested that ground filtering remained challenging in surfaceswith rough terrain or discontinuous slope, dense forest canopies and regions with low vegetation.With improved Lidar systems and growing availability of Lidar data, research on Lidar-basedDTM generation is receiving increasing attention. Although some review papers [1,5–7] have reviewedSensors 2017, 17, 150 3 of 24a large body of ground filtering algorithms, these papers did not include the latest developments ofLidar-based DTM generation, which effectively addressed some critical issues proposed by previousstudies and improved the reliability of general DTM generation. To this end, this study aims to conducta comprehensive review of existing DTM generation methods and provides useful information forscholars to choose, implement and improve DTM generation methods. Firstly, the general principle andnecessary data processing of DTM generation is introduced. Following this, existing DTM generationmethods are classified into different categories and a brief explanation is given to each approach. Next,we summarize some improvement and main challenges in the subject of DTM generation, based onthe understanding of existing methods. Furthermore, we propose some potential solutions to thesechallenges and suggest some promising directions for future studies.2. General Principle of Digital Terrain Models (DTM) GenerationAlthough DTM generation methods vary, most of them share several main steps (albeit conductedin different orders): data pre-processing, ground point filtering and interpolation (except for TIN-basedalgorithms). A brief explanation of these tasks is given as follows.2.1. Data Pre-ProcessingTraditionally, Lidar systems can be classified as discrete Lidar and full-waveform Lidar. Differentfrom discrete Lidar systems, which can usually record multiple echoes per emitted pulse, thefull-waveform Lidar systems can record the complete waveform of backscattered signal echoes.Compared with discrete Lidar data, full-waveform Lidar data provides additional attributes, such asthe neighborhood relationships between waveforms [8] and the differential laser cross-section [9,10] forclassification in forests and urban areas. Although some researchers [11,12] employed full-waveformLidar data for DTM generation, most studies simply decomposed full-waveform Lidar data intodiscrete Lidar point clouds and improves the accuracy of output DTMs by producing dense points.In this case, this paper mainly introduces the pre-processing of discrete Lidar data.Pre-processing is necessary before raw Lidar point clouds can be applied to DTM generation.The most important task for Lidar data pre-processing is the removal of outliers. Due to the existenceof ray multipath and Lidar system error, some points from the Lidar point cloud are of extremely lowelevation. According to their elevation, there are two types of outliers: global outliers and local outliers.Global outliers have elevation values that are the lowest across the study site. The elevation valuesof local outliers are obviously lower than their surrounding points, but not always the lowest in thestudy site. Most DTM generation methods are based on morphological assumption. The lowest pointin each cell is usually regarded as a ground point. If an outlier with extremely low elevation is set as aground point, serious biases will be produced in the process of ground point filtering. Furthermore,this type of bias can hardly be corrected. In this case, effective removal of outliers is a key step forDTM generation.The elevation difference between global outliers and local outliers requires different filteringstrategy. The global outliers are the lowest points in the data set and take up a small proportion ofall points. Therefore, it is comparatively easy to filter these noises. Quantile classification can beconducted and each point is classified into different groups according to its elevation value. Next,researchers can remove a small proportion of (in accordance with the quality of data sets) obviousglobal outliers from the raw data [13–15]. Thus, the point density of the data set is not reduced muchand most obvious noises can be filtered.Local outliers are much lower than neighboring points, but are not always global lowest points.As a result, simple statistical analysis cannot detect and filter local outliers. Some algorithmshave been proposed for detecting local outliers. Outliers can be filtered using distribution-basedapproaches [13,14], mathematical morphology methods [16,17], the density-based method [18,19] andextended local-minimum method [2]. Silven-Cardenas and Wang [14] and Meng et al. [13] employedthe elevation histogram to remove points with extreme elevation values and a Delaunay triangulationSensors 2017, 17, 150 4 of 24to examine remaining outliers. For each point, Kobler et al. [17] computed the vertical differenceD between its elevation and the mean elevation of its neighbouring points. Next, all points wereranked in accordance with D. Following this, P percent of points with the largest negative D value wasdiscarded as outliers. Chen et al. [16] assumed that outliers were scattered and meters lower than theirneighbouring points. Therefore, those points that were obviously lower than their neighbouring pointswere deleted as outliers. Breunig et al. [18] proposed a local outlier factor (LOF) to measure the degreeto which one object is isolated from its surrounding neighborhood. The object that is highly isolatedfrom its neighboring objects is more likely to be a local outlier. By setting appropriate thresholds forthe LOF value, different outliers can be efficiently detected. Those points with a large LOF value wereremoved as outliers. To filter outliers that may appear in various scales, Sotoodeh [19] employedglobal and then local outlier detection using Delaunay triangulation, Euclidean Minimum SpanningTree (EMST) generation and Gabriel Graph (GG) generation. Chen et al. [2] analyzed the elevation ofseveral lowest points in each cell to decide the appropriate ground point. If the difference betweenthe elevation of the lowest point and the mean elevation of rest lowest points is smaller than a giventhreshold, then the lowest point can be set as the ground point. If the difference is larger than thethreshold, then the lowest point is more likely an outlier and thus discarded. Then the second lowestpoint is examined following the same process. This iteration continues until one qualified groundpoint is found.In addition to elevation outliers, some methods employ the intensity for DTM generation.Therefore, not only elevation outliers, but also intensity outliers can lead to serious biases in theprocess of DTM generation. Many factors can negatively influence the intensity value of Lidarpulses. Vain and Kaasalainen [20] introduced different factors that may affect the intensity and themethodology for intensity correction. When elevation and intensity outliers have been filtered, thepoint cloud is ready for ground point filtering.2.2. Ground Point FilteringFiltering ground points from raw point clouds is the crucial step for DTM generation. Although alarge body of studies has proposed different ground filtering algorithms, some general strategies arecommonly employed. Firstly, some seed ground points are selected. As introduced above, most studiesregard the lowest point within a cell as the ground point. Following this, rest points are classified asground points and non-ground points by analyzing the spatial correlation between unclassified pointsand pre-set ground points. As we know, terrain surface is continuous. The closer two ground pointsare, the smaller their elevation difference is. In other words, closely located ground points usuallyhave a similar elevation value whilst non-ground points may cause sudden relief change across a shorthorizontal distance. Following this rule, researchers can set slope (elevation difference/horizontaldistance) thresholds and employ the pre-set ground points to examine the remaining unclassifiedpoints using these thresholds. If the slope between a candidate point and the seed ground point issmaller than the threshold, then the elevation difference between them is more likely to be caused byterrain relief and the candidate point is set as a ground point. If the slope is larger than the threshold,then the elevation difference is more likely to be caused by non-ground objects and this candidate pointis set as a non-ground point. This rule is simplified and ideal, and has been extended and versifiedby many scholars. Nevertheless, the use of morphological thresholds is a fundamental strategy forground point filtering using Lidar point clouds.Methodologies of ground point filtering are the main focus of this paper and the introduction ofspecific algorithms will be presented in the following section.2.3. InterpolationDTMs may be presented as raster images or triangulated irregular network (TIN). Excepted forTIN-based DTMs, interpolation is required to transfer scattered ground points to grid-based DTMs.There are two main types of interpolation methods: the deterministic and probabilistic methods [21].Sensors 2017, 17, 150 5 of 24Deterministic methods regard the estimated value of un-sampled areas as the true value without anyuncertainty. This type of method is effective when sampling points are densely distributed and thephysical mechanic is known. When sampling points are sparsely distributed or the physical mechanicis unknown, it is inappropriate to ignore the estimation error. Inverse distance weighted (IDW), trendsurface (TS) and radial basis function (RBF) are commonly used deterministic methods. IDW and TSfail to feature mountain ridges and valleys well when Lidar data is not dense enough. RBF enablesresearchers to retain such terrain features as mountain ridges and valleys if parameters of RBF areset properly. However, for a large study site, it is very difficult for researchers to employ unifiedparameters to build complicated terrain surface.Probabilistic methods hypothesize that there is a random variable for each location and a setof fixed values for an unsampled location. When an estimated value is given to the unsampledlocation, the probability of occurrence can be calculated as well [21]. This type of method worksefficiently when prior knowledge is missing or the point density is low. Linear prediction, Krigingand conditional simulation are typical probabilistic methods. These methods employ the variogramto estimate the missing value [22]. Through interpolation, the spatial variation of the study areais comprehensively considered. Therefore, this type of method is more likely to be generalized.Furthermore, the uncertainty feature of estimated value gives researchers significant reference on thereliability of results. However, the Kriging and linear prediction method (mathematically equivalentto Kriging method) may lead to smoothing effects and the loss of some terrain details whilst theconditional simulation may result in large estimation errors.Although experiments have been carried out to examine the performance of different interpolationmethods, Fisher and Tate [23] pointed out that there seemed to be no preferable interpolation algorithmfor Lidar data for all landscapes. As a result, researchers are suggested to choose appropriateinterpolation algorithms according to specific terrain situation. If a large study site is of complicatedterrain relief and landscape features, researchers can divide the entire data set into small parts accordingto terrain characteristics and conduct interpolation respectively.3. DTM Generation Methods for Different CategoriesAs explained above, ground point filtering is the key factor for DTM generation. In the pastdecades, ground point filters have been massively studied. Since each method may be understoodfrom different perspectives, the classification of ground point filters is flexible and there are no fixedcategories. Inspired by previous studies [1,5–7] and recent development of new methods, we classifyexisting ground point filters into a set of categories according to their filtering strategies. The generalintroduction and typical methods for each filter category are introduced as follows.3.1. Surface-Based AdjustmentThe use of local minima is one of widely used ground filtering methods. This method regardsthe lowest point in a moving window as a ground point. By moving this window, one ground pointcan be selected for each cell and a DTM can thus be established using these local-minimum points.This method works effectively in flat terrain with few non-ground objects but struggles to achieve thebalance between fine resolution (requiring a small grid size) and few noises (requiring a large grid sizeto remove large non-ground features) in such landscapes as urban areas and dense forests.To solve this problem, researchers proposed different approaches. Amongst these methods,surface-based adjustment has become one frequently employed strategy for generating high qualityDTMs. This type of method first creates an initial surface using part of the control ground points. Then,according to different error characteristics (e.g., the residual to the initial surface), qualified groundpoints are filtered and added to the initial surface for refining the output DTM. Through iterations, theinitial surface is gradually adjusted to a fine-resolution DTM with satisfactory accuracy.Kraus and Pfeifer [24,25] proposed a DTM generation algorithm based on robust linear prediction,which has been widely accepted by researchers. Firstly, a rough estimation of the terrain surfaceSensors 2017, 17, 150 6 of 24is computed using some control ground points (usually acquired using the lowest point for eachcell). Next, the residuals (oriented distances from the surface to measured points) are calculated andeach point is then given a weight according to its residual. Points with a high weight attract thesurface whilst points with a low weight have limited influence on the run of the surface. The iterationcontinues until a stable surface is acquired or the maximum number of iteration is reached. Inspiredby this algorithm, some other methods have also been presented following the principle of refining theDTM gradually [26–28]. Pfeifer et al. [27] and Wack and Wimmer [28] converted the original Lidarpoint clouds to raster images and conducted hierarchical calculation that significantly enhanced thefiltering efficiency. By comparing the coarse DTM with high-resolution sources, the output DTMwas improved gradually. Elmqvist [26] employed an active shape model to approximate the realterrain surface. By minimizing the energy function through iteration, this method is effective for DTMgeneration using very dense point clouds. Kobler et al. [17] proposed a repetitive interpolation method.This approach attempts to filter non-ground points through several steps. First, some traditionalmethods are employed to preliminarily remove most outliers and non-ground returns. Followingthis, a REIN (REpetitive INterpolation) strategy is proposed to further filter remaining non-groundpoints. The REIN method generates a series of TINs using randomly selected sets of Lidar points andthen calculates the distribution of estimate elevation at different DTM locations. By comparing thedistribution of estimate elevation with the global mean offset, remaining non-ground points can beeffectively filtered. Chen et al. [29] used an iterative terrain recovery approach for DTM generation.This algorithm first registers last-return points, and then layers them by dividing Lidar points intodifferent elevation layers. Then the detection of ground points and refinement of the output DTMis conducted from the top layer to the bottom layer. Recently, some advanced surface-filters wereproposed to further improve the reliability of DTM generation. Maguya et al. [30] proposed an adaptivealgorithm for large-scale DTM interpolation in steep forested terrain. Firstly, a set of local minimapoints were selected to produce trend surface for the following process. Different from previousstudies, this method employed two different equations to simulate both linear and quadratic trendsurfaces. A candidate point in the cloud was examined by the trend surface and considered as a ground(non-ground points) if the updated trend surface with this candidate point was of r2 larger (smaller)than the original surface. Through iteration, two DTMs generated using the linear and the quadraticmodels. The DTM with smaller r2 was discarded. If neither model achieved satisfactory result, a cubicspline model was employed to generate DTMs under steep situations. Through comprehensive use ofmultiple surface filters, this method proved to be effective even in some steep areas. Zhang et al. [31]brought a cloth simulation filter (CSF), which is a recent development of computer science, into DTMgeneration. First, the point cloud is turned upside-down, and then a cloth (originally flat) is put on theinverted surface. Next, the shape of the cloth (position of particles) was adjusted through functionsthat explained gravity, intersections and inner forces on the cloth. Finally, the types of cloth particles(unmovable and movable) were used to examine Lidar points in the cloud and qualified groundpoints were filtered. This method further employed a post-processing method to deal with sharplychanging terrain. The results proved that the CSF achieved satisfactory accuracy in most terrain andthe post-processing significantly enhances the performance of CSF in steep terrain.A schematic diagram to explain the principle of surface-based DTM generation methods isdemonstrated as Figure 1.The surface-based filters were widely used and achieved satisfactory accuracy in most terrainsituations. However, this type of method may be problematic in preserving terrain details (e.g., sharpridges and cliffs) and tend to misclassify small non-ground objects [32].Sensors 2017, 17, 150 7 of 24(unmovable and movable) were used to examine Lidar points in the cloud and qualified groundpoints were filtered. This method further employed a post-processing method to deal with sharplychanging terrain. The results proved that the CSF achieved satisfactory accuracy in most terrain andthe post-processing significantly enhances the performance of CSF in steep terrain.A schematic diagram to explain the principle of surface-based DTM generation methods isdemonstrated as Figure 1.Figure 1. The schematic diagram of surface-based DTM generation methods. (a) A lowest point isselected for each cell; (b) A coarse surface is produced based on these pre-selected points; (c) Arefined DTM is generated based on the residue between the coarse surface and the elevation of restpoints.Figure 1. The schematic diagram of surface-based DTM generation methods. (a) A lowest point isselected for each cell; (b) A coarse surface is produced based on these pre-selected points; (c) A refinedDTM is generated based on the residue between the coarse surface and the elevation of rest points.3.2. Morphology-Based FilteringIn addition to surface-based adjustment, many researchers have proposed methods for generatingDTMs based on morphological filtering. This type of method generally works by designing specificslope operators, which describe admissible elevation differences depending on horizontal distances.Vosselman [33] proposed a filtering approach based on mathematical morphology and retained terraindetails by analyzing elevation differences among neighboring points. This method was varied bysome scholars [34–36]. To reduce the influence of terrain relief, Sithole [34] introduced a local operatorthat can alter parameters as a function of the slope of the terrain, whilst Roggero [35] consideredlocal morphology by setting the value of terrain parameters. Zakšek and Pfeifer [36] also proposedan inclined slope operator to follow the terrain. They further pointed out that elevation differencesdownwards are of different characteristics compared with the upwards elevation differences, whichhas not been discussed much by previous studies. Shao and Chen [37] proposed a “climbing andsliding” method for ground point filtering. By emulating natural movements of climbing and sliding,this method performs a local search whilst preserving advantages of a global treatment. Furthermore,some additional geometric features [38], such as erosion and dilation, as well as slope operators, havebeen employed to understand sloped terrain elevations. Some researchers further take into accountlocal context for better generating DTMs. Lu et al. [3] proposed a hybrid conditional random field forautomatic DTM generation. This method applies supervised learning techniques to classify groundand non-ground points and contains both discrete and continuous latent variables. For points classifiedas ground, the LiDAR measurements are used as an estimate of the terrain elevation. Additionally, fornon-ground points, a Gaussian random field is employed for approximating the terrain elevation atthese spots using nearby values.The scan-line strategy is also an important morphology-based DTM generation method. This typeof algorithm aims to detect ground points according to elevation or slope profiles [13,39]. Shan andSampath [39] filtered urban ground points by applying a one-dimensional (1D) and bi-directionallabelling filter, whereas Meng et al. [13] further extended the bi-directional filter to a multi-directionalfiltering method. Wang and Tseng [40] divided traditional a 1D cone-shaped slope operator into twoseparate operators and then employed the two operators to filter ground points respectively. The finalset of grounds was the union of filtered ground points using each operator. The dual-directional profilefilter (DS) filters can even be employed in both vertical and horizontal directions. The DS filter iscapable of detecting sharply changing terrain, even manually made stairs, and thus suitable for urbanlandscapes. Hu et al. [41] employed two strategies for better removing non-ground objects withoutlosing terrain relief. One strategy was to adapt slope operator according to terrain saliency. Throughground point segmentation, the value of terrain saliency was decided by the height difference betweenneighboring segments. Another strategy was to filter candidate points based on line-scanning fromeight directions, instead of one direction. Since the computation involved in this method was simple,the Semi-Global Filter produced reliable DTMs with high efficiency.Sensors 2017, 17, 150 8 of 24Recently, growing emphasis has been given on the design of advanced morphological operators.Li et al. [42] proposed an improved Top-Hat filter for ground point filtering. This main improvementof this new Top-Hat filter was the use of additional sloped brims along with the traditional Top-Hatoperator. In this case, the specific shape of this new operator worked efficiently to distinguish theelevation caused by non-ground objects (especially large buildings) and terrain relief; this operatoris thus highly suitable for urban areas. Susaki [43] proposed an adaptive morphological operatorto reserve local terrain details. The original plane surface for each cell was created by fitting aspecific planar equation and the optimal plan surface was achieved by minimizing the root meansquare error (RMSE). For each iteration, the maximum slope for the current DTM was calculatedas the slope operator to filter new ground points, which were added to update the DTM. Then, theslope operator was also adapted according to the updated DTM. Through this strategy, local terraincan be reserved effectively and automatically. Pingel et al. [44] proposed a simple morphologicalfilter (SMRF). This method works by integrating a linearly increasing window with simple slopethresholding. SMRF not only works as an independent morphological filter, but also serves as astable foundation, based on which other advanced progressive filters may be established. Mongusand Zalik [32] employed some connected operators for better filtering ground points. First, a gridwas produced to establish the connections between points. Secondly, outliers were removed usingstructuring elements. Following this, some connected operators, including the area of the largestcontained object, the maximal roughness of the contained objects, and the level difference by which anon-ground object should be above the neighborhood in order to be recognized, were employed tofilter ground points. The use of multiple thresholds at different scales enables the method to maintainlarge terrain relief (e.g., mountain peaks) without retaining a variety of non-ground objects. Althoughthis method may be problematic in removing attached objects, which is a common difficulty for mostalgorithms, experimental results proved that this filter worked in the most challenging areas with highcomputational efficiency.A simplified demonstration of morphology-based DTM generation method is shown as Figure 2.Sensors 2017, 17, 150 8 of 24update the DTM. Then, the slope operator was also adapted according to the updated DTM.Through this strategy, local terrain can be reserved effectively and automatically. Pingel et al. [44]proposed a simple morphological filter (SMRF). This method works by integrating a linearlyincreasing window with simple slope thresholding. SMRF not only works as an independentmorphological filter, but also serves as a stable foundation, based on which other advancedprogressive filters may be established. Mongus and Zalik [32] employed some connected operatorsfor better filtering ground points. First, a grid was produced to establish the connections betweenpoints. Secondly, outliers were removed using structuring elements. Following this, some connectedoperators, including the area of the largest contained object, the maximal roughness of the containedobjects, and the level difference by which a non-ground object should be above the neighborhood inorder to be recognized, were employed to filter ground points. The use of multiple thresholds atdifferent scales enables the method to maintain large terrain relief (e.g., mountain peaks) withoutretaining a variety of non-ground objects. Although this method may be problematic in removingattached objects, which is a common difficulty for most algorithms, experimental results proved thatthis filter worked in the most challenging areas with high computational efficiency.A simplified demonstration of morphology-based DTM generation method is shown as Figure 2.Figure 2. The schematic diagram of morphology-based DTM generation method. If the slopebetween the ground point and a candidate point is smaller than a (global or local) slope threshold,then this candidate point is set as a ground point. Otherwise, this candidate point is set as anon-ground point.Compared with surface-based filters, Mongus and Žalik [32] suggested that morphology-basedfilters were fairly robust for steep regions and were capable of removing small non-ground objectsand preserving morphology details. Since the scale of morphological filters decides the filteringefficiency of non-ground objects with different sizes, a proper setting of the structuring element is,therefore, a major challenge for morphological filters [32]. Thus, morphology-based filtering may bechallenging in terrains with a variety of non-ground objects.3.3. Triangulated Irregular Network (TIN)-Based RefinementAnother increasingly employed strategy is to generate DTMs with the help of a triangulatedirregular network (TIN). In general, this type of method establishes a preliminary TIN using localminimum points, and then employs diverse examination methods to include qualified groundpoints and refine the TIN gradually. Axelsson [45] established a TIN with local minimum points andanalyzed the relationship between residual points and the TIN. If a residual point meets certaincriteria, it is included in the TIN to refine it. To avoid the edge-cutting effect, a method of mirrorpoints is used to keep qualified edge points. Following the procedure, all ground points can beadded to the final TINFigure 2. The schematic diagram of morphology-based DTM generation method. If the slope betweenthe ground point and a candidate point is smaller than a (global or local) slope threshold, then thiscandidate point is set as a ground point. Otherwise, this candidate point is set as a non-ground point.Compared with surface-based filters, Mongus and Žalik [32] suggested that morphology-basedfilters were fairly robust for steep regions and were capable of removing small non-ground objects andpreserving morphology details. Since the scale of morphological filters decides the filtering efficiencyof non-ground objects with different sizes, a proper setting of the structuring element is, therefore, amajor challenge for morphological filters [32]. Thus, morphology-based filtering may be challengingin terrains with a variety of non-ground objects.3.3. Triangulated Irregular Network (TIN)-Based RefinementAnother increasingly employed strategy is to generate DTMs with the help of a triangulatedirregular network (TIN). In general, this type of method establishes a preliminary TIN using localminimum points, and then employs diverse examination methods to include qualified ground pointsSensors 2017, 17, 150 9 of 24and refine the TIN gradually. Axelsson [45] established a TIN with local minimum points and analyzedthe relationship between residual points and the TIN. If a residual point meets certain criteria, itis included in the TIN to refine it. To avoid the edge-cutting effect, a method of mirror points isused to keep qualified edge points. Following the procedure, all ground points can be added tothe final TIN. Sohn and Dowman [46] employed a “downward and upward divide-and-conquertriangulation” strategy to refine DTM iteratively. First of all, a coarse TIN surface is established usingsome pre-selected ground points. Then “downward divide-and-conquer” is conducted to find pointslower than the trend surface and update the TIN model using these new ground points. Next, “Upwarddivide-and-conquer” is conducted to examine the spatial relationship between the rest of points andthe TIN model. A hypothesis model is used to find candidate ground points which can be added intothe TIN surface and divide local area into more planar terrain surfaces. When more than one candidatepoints exist for a planar terrain surface, minimum description length criterion (MDL) is used to decidethe most reliable points. This iteration continues until no new ground points can be added into theTIN model. Guan el al. [47] proposed a cross-section-plane (CSP)-based filtering method. First, theentire point cloud was put into 3D grids, and multi-directional CSPs—which featured complicated 3Dobjects using 2D surfaces—were produced for each cell to present the DSM. Secondly, potential groundpoints are selected for each CSP according to the feature of elevation, intensity and multi-returns. Sincethis algorithm was conducted in the forested areas and laser pulses can penetrate leaves, points withsingle return were preliminarily set as ground points and some other potential ground points werefiltered by setting judging rules (e.g., the number of neighboring ground points, elevation or intensitydifferences to neighboring points) based on the pre-set ground points. Next, all those potential groundpoints were further filtered by segmenting the scan line and selecting only the lowest point withinthe cell. Finally, a “merging-or-intersecting” processing is conducted to decide final ground points bymerging and intersecting the sets of ground points acquired using each directional CSP, respectively.Chen et al. [48] pointed out that the ground points on the mountain ridges were more likely tocause problems in the TIN-based methods. Classic TIN filters may sacrifice some terrain relief to avoidthe inclusion of non-ground objects. Chen et al. [48] employed three strategies to solve this problem.First, the concept of ridge triangles and adjacent triangles were introduced to detect ridge ground pointsbased on specific characteristics of the two types of triangles. Secondly, a confidence interval estimationmethod was employed to better select seed ground points, including both local lowest points andextracted ridge points. Finally, a simple equation was used to control the number of iterations and thusthe computation efficiency has been improved significantly. Similarly, to retain terrain details aroundbreak lines, Zhang and Lin [49] combined the progressive TIN densification (PTD) with segmentationusing smoothness constraint (SUSC). Firstly, original seed ground points were selected, as a key step inPTD processing. Next, regional growing method was employed based on original seed ground pointsand more seed ground points were filtered for the following TIN densification. With many more seedgrounds, the discontinuities and terrain relief can be better reserved.A schematic diagram of TIN-based DTM generation method is shown in Figure 3.Sensors 2017, 17, 150 9 of 24area into more planar terrain surfaces. When more than one candidate points exist for a planarterrain surface, minimum description length criterion (MDL) is used to decide the most reliablepoints. This iteration continues until no new ground points can be added into the TIN model. Guanel al. [47] proposed a cross-section-plane (CSP)-based filtering method. First, the entire point cloudwas put into 3D grids, and multi-directional CSPs—which featured complicated 3D objects using 2Dsurfaces—were produced for each cell to present the DSM. Secondly, potential ground points areselected for each CSP according to the feature of elevation, intensity and multi-returns. Since thisalgorithm was conducted in the forested areas and laser pulses can penetrate leaves, points withsingle return were preliminarily set as ground points and some other potential ground points werefiltered by setting judging rules (e.g., the number of neighboring ground points, elevation orintensity differences to neighboring points) based on the pre-set ground points. Next, all thosepotential ground points were further filtered by segmenting the scan line and selecting only thelowest point within the cell. Finally, a “merging-or-intersecting” processing is conducted to decidefinal ground points by merging and intersecting the sets of ground points acquired using eachdirectional CSP, respectively.Chen et al. [48] pointed out that the ground points on the mountain ridges were more likely tocause problems in the TIN-based methods. Classic TIN filters may sacrifice some terrain relief toavoid the inclusion of non-ground objects. Chen et al. [48] employed three strategies to solve thisproblem. First, the concept of ridge triangles and adjacent triangles were introduced to detect ridgeground points based on specific characteristics of the two types of triangles. Secondly, a confidenceinterval estimation method was employed to better select seed ground points, including both locallowest points and extracted ridge points. Finally, a simple equation was used to control the numberof iterations and thus the computation efficiency has been improved significantly. Similarly, toretain terrain details around break lines, Zhang and Lin [49] combined the progressive TINdensification (PTD) with segmentation using smoothness constraint (SUSC). Firstly, original seedground points were selected, as a key step in PTD processing. Next, regional growing method wasemployed based on original seed ground points and more seed ground points were filtered for thefollowing TIN densification. With many more seed grounds, the discontinuities and terrain reliefcan be better reserved.A schematic diagram of TIN-based DTM generation method is shown in Figure 3.Figure 3. The schematic diagram of morphology-based DTM generation method. (a) Some of thelowest points are selected as preliminary ground points and form a coarse TIN; (b) Rest points areexamined using triangles within this TIN model.The most widely used Lidar processing software TerraScan was designed based on theAxelsson’s TIN-model [45], and the reliability and accuracy of this method has been proved by alarge body of studies. However, Chen et al. [48] pointed out that progressive TIN densification(PTD) filters may have difficulties in detecting discontinuous terrains, such as sharp ridges. Thussome targeted processing should be conducted for better results. Furthermore, PTD is generally timeconsuming due to numerous implementations of TIN construction for a large number of points [48].Figure 3. The schematic diagram of morphology-based DTM generation method. (a) Some of thelowest points are selected as preliminary ground points and form a coarse TIN; (b) Rest points areexamined using triangles within this TIN model.Sensors 2017, 17, 150 10 of 24The most widely used Lidar processing software TerraScan was designed based on the Axelsson’sTIN-model [45], and the reliability and accuracy of this method has been proved by a large bodyof studies. However, Chen et al. [48] pointed out that progressive TIN densification (PTD) filtersmay have difficulties in detecting discontinuous terrains, such as sharp ridges. Thus some targetedprocessing should be conducted for better results. Furthermore, PTD is generally time consuming dueto numerous implementations of TIN construction for a large number of points [48].3.4. Segmentation and ClassificationWith the growing applications of airborne Lidar data in different areas, methodologies for imageprocessing and land use (cover) classification provide important reference for DTM generation. Byputting unclassified points from raw Lidar point clouds into different classes according to classificationrules [50–58], filtered ground points can then be used for establishing DTMs. Theoretically, availablefeatures from Lidar data for land cover classification are only elevation and one band of intensity(as compared to tens, even hundreds of bands in multi-spectral remote sensing). With limitedattributes, it is very difficult to conduct point-based classification using Lidar point cloud. To fillthis gap, some scholars employed additional features, such as the number of returns [59,60] or theelevation difference between the first and the last return [61–63] to separate ground points fromnon-ground points. Furthermore, growing research emphasis has been given to the object-basedmethod for DTM generation. Owing to the high resolution of Lidar data, neighbouring points arehighly correlative in terms of both elevation and intensity, which makes Lidar data suitable forobject-based classification [64,65]. This type of segmentation and classification-based DTM generationmethod is usually conducted in several steps. First, the raw point cloud is interpolated to raster images.Considering the materials for hierarchical image segmentation, not only the feature of elevation, butalso features of intensity and elevation difference are converted to images. Next, image segmentationis conducted by setting some segmentation rules (e.g., segmentation scale, weight for each image layer,compactness.) and the entire image is segmented into unclassified objects. By setting classificationrules for each land use (cover) type in terms of elevation, intensity, elevation difference and somegeometric attributes (e.g., area, perimeter, shape index, roundness, etc.), different non-ground featurescan be filtered effectively and DTMs can thus be generated. Antonarakis et al. [66] employed vegetationheight models, percentage canopy hit models, intensity models and skewness and kurtosis modelsto classify Lidar points in forest areas and bare-ground and other non-ground land cover types wereseparated with high accuracy. Johansen et al. [67] monitored the environmental condition of riparianzones by assessing some riparian condition indicators and efficiently classified riparian vegetationand ground by conducting object-based image analysis. Im et al. [68], Samadzadegan et al. [69]and Huang et al. [70] employed a diversity of gray-level co-occurrence Matrix (GLCM) textures (e.g.,homogeneity, mean, entropy, correlation and dissimilarity) to classify trees, buildings and ground andachieved satisfactory accuracy.Recently, some researchers proposed some new rules to improve the segmentation andclassification-based filters. Niemeyer et al. [54] integrated a Random Forest classifier into a conditionalrandom field (CRF) framework to classify grassland, road, different types of buildings, trees andlow vegetation. The texture features, which have been widely used in processing remote sensingimages, are also important tools for DTM generation using Lidar data. Chen et al. [71] conductedimage-segmentation and then employed a shortest-distance-pair strategy, as well as a set of slopeoperators, to classify ground segments, instead of ground points. Thus this method is specificallysuitable for removing a variety of urban objects. After image segmentation, Zhang et al. [72] selected13 features of the geometry, radiometry, topology and echo characteristics for filtering groundsegments, and the SVM tool was employed for better feature selection. Additionally, an improvedconnected-component labeling method, which aims to remove small and isolated segments, wereemployed for a more homogeneous classification results.Sensors 2017, 17, 150 11 of 24A schematic diagram of segmentation and classification-based DTM generation methods is shownSensors as Figure2017,417 ., 150 11 of 24Figure 4. The schematic diagram of segmentation and classification-based DTM generation method.(a) Raw Lidar points; (b) Raster images produced using raw Lidar points; (c) After-imagesegmentation, unclassified segments are categorized into different land cover types by employing adiversity of features.The segmentation and classification-based methods make full use of additional geometric,texture and other features for better filtering ground points, and producing more homogeneousDTMs. This type of method can effectively remove non-ground objects of different sizes and shapes,and is thus highly suitable for urban areas. Nevertheless, Chen et al. [71] pointed out thesegmentation-based DTM generation methods may struggle in densely forested areas, as scatteredlaser pulses that pass the gaps of tree crowns and reach the ground cannot satisfy the requirement ofimage segmentation. Furthermore, the reliability of this method greatly depends on the accuracy ofsegmentation and thus tuning segmentation parameters causes uncertainty to the performance ofthis type of method.3.5. Statistical AnalysisIn recent years, statistical analysis for DTM generation has attracted growing attention. Bartelsand Wei [73] proposed an object feature extractor for terrain feature extraction. Firstly, the originalLidar point cloud was regularly gridded. Following this, the resulting matrix was decomposed andanalyzed using wavelets. The result proved that detached objects could be detected in challenginghilly terrain. However, this method performs poorly in detecting large flat roofs and structures suchas bridges. Bretar and Chehata [74] employed a Bayesian method for regularizing initially acquiredDTM. This approach was based on the definition of an energy function that managed the refinementof a terrain surface and the process of minimizing this energy led to the final DTM.Some exciting progress in this field is the emerging threshold-free algorithms for DTMgeneration. Bartels and Wei [75] proposed an unsupervised Lidar filtering algorithm, skewnessbalancing. This method simply considers the skewness value of the remaining point cloud anddeletes the highest point in the point cloud if the skewness value is larger than 0. As a result, verylimited human interaction is required and this method is thus highly automatic. Yao et al. [76] andBao et al. [77] also employed skewness balancing and further developed this algorithm byintegrating skewness value with other classification features. Mongus and Zalik [78] proposed anunsupervised, parameter-free DTM generation method. Firstly, traditional morphological opening(with a large window size) and closing (with a small window size) was conducted repeatedly toremove lower outliers. Secondly, a series of control points to producing a pyramidal hierarchicalstructure for the multi-scale filtering. The selection of control points was done by employing anautomatic bottom-up strategy. Hence, the selection of upper-layer control points was decided bypre-selected control points within the four neighboring cells in the lower scale. Next, thin platespline (TPS) interpolation was conducted to produce smooth, oscillation-free trend surfaces. Finally,the ground point was filtered by analyzing the residual between the elevation of the candidate pointand the trend surface. Different from previous studies, there was no pre-set threshold for theresidual to filter ground points. Instead, the mean residual of all remaining points was calculatedautomatically and the ground point filtering was done accordingly. Additionally, the DTMresolution, which was required by multi-scale filtering, was updated automatically. As a result, thismethod can be highly effective and automatic. The experimental results proved this algorithmachieved high accuracy even in complicated terrains.Figure 4. The schematic diagram of segmentation and classification-based DTM generation method. (a)Raw Lidar points; (b) Raster images produced using raw Lidar points; (c) After-image segmentation,unclassified segments are categorized into different land cover types by employing a diversityof features.The segmentation and classification-based methods make full use of additional geometric, textureand other features for better filtering ground points, and producing more homogeneous DTMs. Thistype of method can effectively remove non-ground objects of different sizes and shapes, and is thushighly suitable for urban areas. Nevertheless, Chen et al. [71] pointed out the segmentation-basedDTM generation methods may struggle in densely forested areas, as scattered laser pulses that passthe gaps of tree crowns and reach the ground cannot satisfy the requirement of image segmentation.Furthermore, the reliability of this method greatly depends on the accuracy of segmentation and thustuning segmentation parameters causes uncertainty to the performance of this type of method.3.5. Statistical AnalysisIn recent years, statistical analysis for DTM generation has attracted growing attention. Bartels andWei [73] proposed an object feature extractor for terrain feature extraction. Firstly, the original Lidarpoint cloud was regularly gridded. Following this, the resulting matrix was decomposed and analyzedusing wavelets. The result proved that detached objects could be detected in challenging hillyterrain. However, this method performs poorly in detecting large flat roofs and structures suchas bridges. Bretar and Chehata [74] employed a Bayesian method for regularizing initially acquiredDTM. This approach was based on the definition of an energy function that managed the refinementof a terrain surface and the process of minimizing this energy led to the final DTM.Some exciting progress in this field is the emerging threshold-free algorithms for DTM generation.Bartels and Wei [75] proposed an unsupervised Lidar filtering algorithm, skewness balancing.This method simply considers the skewness value of the remaining point cloud and deletes thehighest point in the point cloud if the skewness value is larger than 0. As a result, very limited humaninteraction is required and this method is thus highly automatic. Yao et al. [76] and Bao et al. [77] alsoemployed skewness balancing and further developed this algorithm by integrating skewness valuewith other classification features. Mongus and Zalik [78] proposed an unsupervised, parameter-freeDTM generation method. Firstly, traditional morphological opening (with a large window size) andclosing (with a small window size) was conducted repeatedly to remove lower outliers. Secondly, aseries of control points to producing a pyramidal hierarchical structure for the multi-scale filtering.The selection of control points was done by employing an automatic bottom-up strategy. Hence, theselection of upper-layer control points was decided by pre-selected control points within the fourneighboring cells in the lower scale. Next, thin plate spline (TPS) interpolation was conducted toproduce smooth, oscillation-free trend surfaces. Finally, the ground point was filtered by analyzingthe residual between the elevation of the candidate point and the trend surface. Different fromprevious studies, there was no pre-set threshold for the residual to filter ground points. Instead,the mean residual of all remaining points was calculated automatically and the ground pointSensors 2017, 17, 150 12 of 24filtering was done accordingly. Additionally, the DTM resolution, which was required by multi-scalefiltering, was updated automatically. As a result, this method can be highly effective and automatic.The experimental results proved this algorithm achieved high accuracy even in complicated terrains.A schematic diagram of statistical analysis-based DTM generation method is shown as Figure 5.Sensors 2017, 17, 150 12 of 24A schematic diagram of statistical analysis-based DTM generation method is shown as Figure 5.Figure 5. The schematic diagram of statistical analysis-based DTM generation method. (a) Raw Lidarpoints; (b) Global (local) high points filtered through statistical analysis; (c) DTM generated withoutfiltered points.Statistical analysis-based filters, especially parameter-free algorithms, reduce the uncertainty ofmanually tuning parameters and make the transfer of specific methods to other study sites morerobust. Moreover, these methods usually perform well in generally flat terrain without complicatednon-ground objects. Compared with parameter dependent filters, the reliability of this type ofmethod may decrease significantly. Some introduced methods (e.g., [78]), achieved good accuracy inmost terrains. This results from the fact that these methods do not purely rely on statistical analysis,and strategies from other ground point filters are considered as well, which is discussed more in thefollowing section.3.6. Multi-Scale ComparisonIn addition to the above discussed categories, increasing studies employ a multi-scalecomparison strategy for DTM generation. This type of DTM generation method is not usuallyconsidered as an independent category. Previous studies may regard these methods as applicationsof surface, morphology or TIN-based methods. The main reason we separate this category fromother well-accepted categories is explained as follows. Firstly, examining a point at different scales toremove noise from a range of non-ground objects is one of its theoretical differences to surface,morphological or TIN-based methods. Secondly, this type of method provides practical and reliablesolutions for integrating merits of DTMs generated using different methods or parameters, which isdiscussed more in Section 4.In general, this type of method works in the following steps. Several preliminary trend surfacesof different resolutions are produced. Each point in the point cloud is then examined at differentscales by comparing the elevation difference between the point and different trend surfaces. If thecandidate point is classified as non-ground points at a small scale (e.g., 2 m), it is assigned as anon-ground point definitely. If it is classified as a ground point at a small scale, it should be furtherexamined at a middle scale (e.g., 10 m). This is because small non-ground objects (e.g., trees), whichwill be filtered at a middle scale, may be retained in the trend surface acquired at a small filteringscale. By analogy, if the point is as well classified as a ground point at a middle scale, it needs to beexamined at a large scale (e.g., 50 m). This is because large non-ground objects (e.g., buildings),which will be filtered at a large scale, may be retained in the trend surface acquired at a middlefiltering scale. If this point continues to be classified as a ground point at a large scale, this point isassigned as a ground point. The key issue for this type of method is the setting of a series of filteringscales. If the scale is too large, then too few ground points can be extracted and the output DTM canbe over-smoothing and most terrain details are lost. On the other hand, if the scale is not largeenough to filter large buildings, many building traces may be retained in the final DTM.Zhang et al. [79] filtered non-ground points using gradually increased window size and a slopeoperator that was decided automatically by comparing the filtered and unfiltered data iteratively.Chen et al. [48] employed increased window size and a building mask to iteratively update theFigure 5. The schematic diagram of statistical analysis-based DTM generation method. (a) Raw Lidarpoints; (b) Global (local) high points filtered through statistical analysis; (c) DTM generated withoutfiltered points.Statistical analysis-based filters, especially parameter-free algorithms, reduce the uncertaintyof manually tuning parameters and make the transfer of specific methods to other study sites morerobust. Moreover, these methods usually perform well in generally flat terrain without complicatednon-ground objects. Compared with parameter dependent filters, the reliability of this type of methodmay decrease significantly. Some introduced methods (e.g., [78]), achieved good accuracy in mostterrains. This results from the fact that these methods do not purely rely on statistical analysis,and strategies from other ground point filters are considered as well, which is discussed more in thefollowing section.3.6. Multi-Scale ComparisonIn addition to the above discussed categories, increasing studies employ a multi-scale comparisonstrategy for DTM generation. This type of DTM generation method is not usually considered asan independent category. Previous studies may regard these methods as applications of surface,morphology or TIN-based methods. The main reason we separate this category from otherwell-accepted categories is explained as follows. Firstly, examining a point at different scales toremove noise from a range of non-ground objects is one of its theoretical differences to surface,morphological or TIN-based methods. Secondly, this type of method provides practical and reliablesolutions for integrating merits of DTMs generated using different methods or parameters, which isdiscussed more in Section 4.In general, this type of method works in the following steps. Several preliminary trend surfacesof different resolutions are produced. Each point in the point cloud is then examined at different scalesby comparing the elevation difference between the point and different trend surfaces. If the candidatepoint is classified as non-ground points at a small scale (e.g., 2 m), it is assigned as a non-ground pointdefinitely. If it is classified as a ground point at a small scale, it should be further examined at a middlescale (e.g., 10 m). This is because small non-ground objects (e.g., trees), which will be filtered at amiddle scale, may be retained in the trend surface acquired at a small filtering scale. By analogy, if thepoint is as well classified as a ground point at a middle scale, it needs to be examined at a large scale(e.g., 50 m). This is because large non-ground objects (e.g., buildings), which will be filtered at a largescale, may be retained in the trend surface acquired at a middle filtering scale. If this point continues tobe classified as a ground point at a large scale, this point is assigned as a ground point. The key issuefor this type of method is the setting of a series of filtering scales. If the scale is too large, then too fewground points can be extracted and the output DTM can be over-smoothing and most terrain detailsSensors 2017, 17, 150 13 of 24are lost. On the other hand, if the scale is not large enough to filter large buildings, many buildingtraces may be retained in the final DTM.Zhang et al. [79] filtered non-ground points using gradually increased window size and a slopeoperator that was decided automatically by comparing the filtered and unfiltered data iteratively.Chen et al. [48] employed increased window size and a building mask to iteratively update theelevation of each point in the point cloud. Li et al. [80] proposed a multi-scale mathematic morphology.First, the method extracts edge points of non-ground objects, and then processes opening operation ofmathematical morphology using remaining points in the local region. Next, by comparing adjacentfilter surfaces with given thresholds, surface features are extracted through increasing window size.Xiong et al. [81] designed a method for the automatic generation of DTMs. First, some regular gridsare generated and non-ground points on each grid are filtered. Next, by changing (either increasingor decreasing) the grid size, new grids are produced and non-ground points retained on these gridsare filtered. The iteration continues until all points have been examined and ground points havebeen recorded for the following DTM generation. Chen et al. [2] presented an upward-fusion DTMgeneration method. Different from many other methods, this method is based on raster calculation,instead of point-based filtering. First, several preliminary DTMs of different grid sizes are producedusing the local minimum method. Next, upward fusion is conducted between these DTMs. This processbegins with a DTM of the largest grid size and a finer-scale DTM is compared with this large-scale DTM.By setting proper thresholds, a new DTM is achieved by acquiring qualified elevation values fromthe finer DTM and keeping the value from the large-scale DTM when the value from the finer DTMis beyond the threshold. Iteration continues until all preliminary DTMs have been processed and arefined DTM of high resolution is generated. The experiment proved that this method produces DTMswith high accuracy. Moreover, upward fusion can be conducted using DTMs generated using differentalgorithms or parameters. Chen et al. [82] presented a multi-resolution hierarchical classification(MHC) algorithm for differentiating ground from non-ground points. MHC includes three levels ofhierarchy, which is featured with the simultaneous increase of window size and residual threshold.For each level, the surface is iteratively approximated to the ground using thin plate spline (TPS) untilno ground point is detected. Following this, these classified ground points are used to update thesurface in the next iteration. Mongus et al. [83] proposed a ground point filter for urban areas. Insteadof several different window sizes, this algorithm employed a function to examine a candidate pointby comparing the elevation difference between this point and its neighboring points within a seriesof continuous window sizes (from 1 to the size of largest building size at this site). By analyzing thecurve shape of this function for each point, ground points were filtered effectively. This algorithm washighly suitable for ground point filtering in urban areas with large flat-roof buildings and a diversityof non-ground objects.Similar to Chen et al. [2]’s upward-fusion algorithm, Maguya et al. [84] first produced a seriesof coarse DTMs and acquired the final DTM by merging these coarse DTMs. The main differencesbetween the two methods were that (1) these coarse DTMs were produced using the surface-basedground point filter [30], instead of the extended local-minimum method; (2) Maguya et al. [83] mergeda series of coarse DTMs at the same time and the merge was conducted by considering the elevationof neighboring cells on coarse DTMs with different resolutions. On the other hand, Chen et al. [2]’salgorithms considered the comparison between neighboring cells before the process of mergingand the upward-fusion was conducted simply in each corresponding cell respectively. Therefore,Maguya et al. [84]’s method may achieve better accuracy, whereas Chen et al. [2]’s method is easier toimplement and can have higher fusion efficiency, which can be very convenient for improving otherDTM generation methods [2]. Su et al. [85] proposed a hierarchical moving curve-fitting algorithm.The initial block size (window size) for filtering ground points was set as the size of the largest buildingin the study site. Then the block size was updated automatically and the filtering parameters wereupdated accordingly. The filtering at each scale was conducted by approximating an optimal trendsurface and then each candidate point was examined by comparing the residual between its elevationSensors 2017, 17, 150 14 of 24and the trend surface. The trend surface was initially simulated using a second-degree polynomialfunction and the parameters of this function for each block was decided by the 16 lowest points in theblock through the least square method. Additionally, to avoid the lack of qualified control points ineach block, this algorithm employed a 25% overlapping strategy during the moving of blocks.A schematic diagram of multi-scale comparison-based DTM generation method is shown asFigure 6.Sensors 2017, 17, 150 14 of 24A schematic diagram of multi-scale comparison-based DTM generation method is shown asFigure 6.Figure 6. The schematic diagram of multi-scale comparison-based DTM generation method.Figure 6 is explained as follows. First, the target point is compared with the lowest point G1 in a2 × 2 cell using a slope threshold λ1 (comparatively small). If the slope between the target point andG1 is larger than λ1, then the target point is more likely from a small non-ground object and thus setas a non-ground point. If the slope is smaller than λ1, then the target point is further compared withthe lowest point G2 in a 4 × 4 cell using a slope threshold λ2 (larger than λ1). By analogy, if the slopebetween the target point and G2 is larger than λ2, then the target point is more likely from a middlenon-ground object and set as a non-ground point. If the slope is smaller than λ2, then the target pointis further compared with the lowest point G3 in an 8 × 8 cell using a slope threshold λ3 (larger thanλ2). If the slope between the target point and G3 is larger than λ3, then the target point is more likelyfrom a middle non-ground object and set as a non-ground point. If the slope is smaller than λ3, thenthe target point is set as a ground point.Multi-scale comparison-based method is of special value for generating DTMs in urban areas,which are characterized by large, flat buildings and a diversity of non-ground features. Comparedwith other methods, the use of a large window size proves particularly effective for filtering largebuildings. By adjusting the window size gradually, urban features of different sizes can be filteredeffectively. However, with limited number of window sizes, this type of method may cause the lossof terrain details and sharply cut terrain relief. On the other hand, a function of automaticallyupdating window sizes can help preserve terrain relief whilst significantly increasing computationtime. In addition, this type of method is more suitable for a generally flat area and may performpoorly in rapidly changing terrain situations.3.7. Overview of Different Ground Point FiltersAs repeatedly stated by a large body of studies, ground point filtering from airborne Lidar dataremains challenging. The main limitation is that ground point filters can have differentperformances under different terrains, particularly complicated terrains. Inspired by previousstudies, the strengths and limitations of each type of filter are summarized in Table 1. As mentioned,the nomenclature and classification of different filtering methods employed in this study divergefrom previous studies. Additionally, the understanding of terrain situations and the suitability offilters may be controversial. Thus, we generally considered the common points from a majority ofscholars. Recently, some advanced ground filtering methods seem capable of solving traditionalchallenges in difficult terrains. However, these filters can be regarded as the combination of methodsthat belong to several different categories. In this section, the comparison between differentcategories was concluded mainly based on those methods that simply depend on one filteringstrategy. Therefore, it provides some references for scholars to select proper filters according tospecific terrains or propose new algorithms based on existing filters.Figure 6. The schematic diagram of multi-scale comparison-based DTM generation method.Figure 6 is explained as follows. First, the target point is compared with the lowest point G1 ina 2 × 2 cell using a slope threshold λ1 (comparatively small). If the slope between the target point andG1 is larger than λ1, then the target point is more likely from a small non-ground object and thus setas a non-ground point. If the slope is smaller than λ1, then the target point is further compared withthe lowest point G2 in a 4 × 4 cell using a slope threshold λ2 (larger than λ1). By analogy, if the slopebetween the target point and G2 is larger than λ2, then the target point is more likely from a middlenon-ground object and set as a non-ground point. If the slope is smaller than λ2, then the target pointis further compared with the lowest point G3 in an 8 × 8 cell using a slope threshold λ3 (larger thanλ2). If the slope between the target point and G3 is larger than λ3, then the target point is more likelyfrom a middle non-ground object and set as a non-ground point. If the slope is smaller than λ3, thenthe target point is set as a ground point.Multi-scale comparison-based method is of special value for generating DTMs in urban areas,which are characterized by large, flat buildings and a diversity of non-ground features. Compared withother methods, the use of a large window size proves particularly effective for filtering large buildings.By adjusting the window size gradually, urban features of different sizes can be filtered effectively.However, with limited number of window sizes, this type of method may cause the loss of terraindetails and sharply cut terrain relief. On the other hand, a function of automatically updating windowsizes can help preserve terrain relief whilst significantly increasing computation time. In addition, thistype of method is more suitable for a generally flat area and may perform poorly in rapidly changingterrain situations.3.7. Overview of Different Ground Point FiltersAs repeatedly stated by a large body of studies, ground point filtering from airborne Lidar dataremains challenging. The main limitation is that ground point filters can have different performancesunder different terrains, particularly complicated terrains. Inspired by previous studies, the strengthsand limitations of each type of filter are summarized in Table 1. As mentioned, the nomenclatureand classification of different filtering methods employed in this study diverge from previous studies.Additionally, the understanding of terrain situations and the suitability of filters may be controversial.Thus, we generally considered the common points from a majority of scholars. Recently, someadvanced ground filtering methods seem capable of solving traditional challenges in difficult terrains.However, these filters can be regarded as the combination of methods that belong to several differentSensors 2017, 17, 150 15 of 24categories. In this section, the comparison between different categories was concluded mainly basedon those methods that simply depend on one filtering strategy. Therefore, it provides some referencesfor scholars to select proper filters according to specific terrains or propose new algorithms based onexisting filters.Table 1. Characteristics of different ground point filters.Filtering Methods Suitable for Not Suitable for Memory Storage Demands Computational Efficiency 2Surface-based Forested areas Rough andsteep terrains High MiddleMorphology-based Steepterrains 1, Terrainswith small objectsTerrains withvarious objects Low HighTIN-based Steep terrains Discontinuous terrains Urban areas, Middle MiddleSegmentation-based Urban areas, Terrainswith various objectsRough and steepterrains, Dense forests NA 3 NA 4Statistical analysis Generally flat terrains Terrains withvarious objects Low Low

Multi-scale comparison Urban areas Rough andsteep terrains Middle Low

1 The suitability of morphology-based filters in steep terrains is controversial. Although some scholars claimedthis type of filter struggles in steep terrains, Mongus and Žalik [32] pointed out that morphological filtersworked fairly well in steep terrains. Due to this controversy, it is suggested that the morphological filtersshould be employed in steep terrains without a diversity of objects, which would otherwise significantlyreduce the reliability of generated DTMs; 2 The computational efficiency of different categories is generalizedaccording to previous experiments [48,49]. Thus, the conclusions made here purely depend on the performanceof some typical algorithms for each category and the time efficiency for specific algorithms can be different;3 The memory storage demands for the segmentation-based method varies significantly with the resolutionof interpolated DSMs and image segmentation scale and thus cannot be compared with other filters; 4 Thecomputational efficiency for the segmentation-based method highly depends on some necessary manualparameter settings during image-segmentation and classification, and can hardly be compared with other highlyautomatic methods.4. Discussion4.1. Recent Progress and Remaining Limitations in DTM GenerationSithole and Vosselman [1] compared the performance of some typical DTM generation methods ina diversity of terrain situations. According to the result of accuracy assessment, algorithms proposedby Axelsson [45] and Pfeifer [27] achieved comparatively better results, as the coarse-to-fine strategyefficiently reduced large biases caused by large terrain relief. However, no method proved effectivefor extracting ground points in complicated terrain scenarios, which were characterized by sharplychanging elevation and densely trees attached to edges.4.1.1. DTM Generation with Improved AccuracySince Sithole and Vosselman [1]’s comparative experiments, many innovative methods have beenproposed, which contribute to a more comprehensive and incisive understanding of ground pointfiltering using airborne Lidar data. First, increasing research on the design of highly automatic andgenerally applicable algorithms has been an exciting progress. Sithole and Vosselman [1] suggestedthat the human editing process took about 60%–80% of the processing time for DTM generation. As aresult, these highly automatic filters, especially threshold-free algorithms [75,78], significantly enhancethe time-efficiency of DTM generation and reduce possible biases caused by human editing. Anothermain challenge for ground point filtering is that the efficiency of most DTM generation algorithms issensitive to specific terrains. In this case, filters that work well in forest (urban) areas may struggle inurban (forest) areas. In the past decade, some researchers [3,32,78] have proposed generally applicableground point filters for a diversity of different terrains, which effectively reduces the time and difficultyin selecting specific DTM generation methods. Finally, based on experiments using the bench dataSensors 2017, 17, 150 16 of 24provided by Sithole and Vosselman [1], the accuracy of DTMs generated using these new algorithmshave been improved notably from previous studies.4.1.2. DTM Generation and Presentation with Improved Computational EfficiencyIn addition to improved accuracy and applicability, some scholars paid special emphasis onhigher computational efficiency for DTM generation. For instance, Mongus and Zalik [32] reducedthe processing time for DTM generation by 98% by arranging the input Lidar data into a Max-Treerepresented grid. The processing time for DTM generation and presentation is not only decidedby the computational efficiency of DTM generation algorithms, but also decided by the memorystorage demands for increasing in data volume. As a result, growing research emphasis has beenput on the design and validation of DTM compression methods [86–90]. Mandlburger et al. [86]proposed an adaptive TIN refinement approach for data thinning and the experiments proved thatthe compression rate for DTMs generated in varying landscapes all exceeded 80%. By predictingthe number of Discrete Cosine Transform (DCT) Coefficients, Forczmanski and Maleika [87] reducedthe processing time for seabed DTM compression by 40%. Based on the predicted number of DCTcoefficients, Forczmanski and Maleika [88] further developed a near-lossless principal componentanalysis (PCA)-based compression algorithm. The PCA-based compression method for generatedseabed DTM demonstrated better accuracy and efficiency than traditional DCT-based methods.Meanwhile, some scholars [89,90] have worked on the better presentation of compressed DTMsin different formats or devices. Quintero et al [89] specifically designed a DTM compression algorithmfor mobile devices. By replacing the full decomposition stage with obtained parameters (e.g., altitudes,contour lines and terrain roughness index) from a sub-region of the DTM, the algorithm achievedan up-to-80% compression rate. Scarmana [90] compared presented DTMs in different compressionformats, including JPEG, WinZip, TIFF and PNG. Due to its high compression rate and short decodingtime, the PNG (Portable Network Graphics) format was an appropriate tool for cross-platform DTMrepresentation, storage, retrieval and display.4.1.3. Remaining Challenges for DTM GenerationIn spite of many new methods for ground point filtering, DTM generation remains challenging.As explained above, each category of methods has its own strengths and limitations. However, fewmethods are applicable to all terrain situations, which include densely forested areas, urban areas,sharp ridges and discontinuous terrains and so forth. Furthermore, it is still difficult to employ onesingle filter for DTM generation in very complicated or highly fragmented terrain situations. In recentyears, airborne Lidar has been increasingly applied to oceanography and hydrology [91]. However,due to the absorption and scattering effects of water on laser pulses, DTM generation in coastal regionshas extra difficulties. Mohammadzadeh and Valadan Zoej [91] concluded some common challengesfor DTM generation in ocean and hydrography, including dune and tidal flat measurement and coastalchange and erosion.4.2. Promising DirectionsAccording to limitations in current research of DTM generation, especially ground point filtering,more emphasis should be given on two directions:4.2.1. Advanced DTM Generation by Combining Different MethodsThis paper introduced different categories of DTM generation methods, and each type of methodis more suitable for a certain terrains]. If the merits of different methods can be combined, theaccuracy of generated DTMs can be improved significantly. As explained, some new ground pointfilters can be generally applied to a variety of terrains and achieve better accuracy than previousalgorithms. The main improvement was that these filters, although classified into specific categoriesin this paper, actually combined different filter strategies. Su et al. [85]’s method successfullySensors 2017, 17, 150 17 of 24combined the surface-based filter and the multi-scale comparison algorithm for better removing largebuildings, as well as preserving terrain details. Mongus and Zalik [32] proposed a DTM generationmethod, which combined the strategy of multi-scale comparison, segmentation and morphologicaloperators, and achieved high accuracy and computational accuracy. Zhang and Lin [49] combinedTIN-based and segmentation-based algorithms for a new filter, which worked effectively both in steepterrains and terrains with a variety of different objects. Mongus et al. [83] combined the strategyof morphological operators, multi-scale comparison and surface-based adjustment and effectivelyclassified ground points from buildings of a diversity of structures. The DTM generation methodproposed by Mongus and Zalik [78] has been frequently used not only because of its parameter-freestrategy, but also due to its general suitability for various terrains. The efficiency of this methodalso mainly depends on the successful combination of the surface-based adjustment, multi-scalecomparison and statistical analysis.In addition to these individual methods that combine different filtering strategies, some scholarseven proposed methods that can directly make improvements on the outputs by merging DTMsgenerated using different methods (software) or parameters, which is a promising tool for significantlyimproving DTM generation based on existing methods. Chen et al. [2] proposed an upwardfusion-based method for DTM generation. This method is a raster-based method, so raster calculationcan be conducted using raster DTMs generated using any methods or parameters. Therefore, thismethod not only works alone to generate high quality DTMs, but also serves as an efficient tool forfusing DTMs generated using different methods. For instance, when upward-fusion was conductedbetween DTMs generated using lasground and TerraScan, five out of nine points with large biaswere filtered and the mean bias of updated DTM was improved from 0.254 m to 0.114 m. Whenupward-fusion was conducted between DTMs generated using TIFFS with different parameter values,16 out of 24 points with large bias were filtered and the mean bias of the updated DTM was improvedfrom 0.755 m to 0.206 m.By designing specific methods combined with different filtering strategies or fusing outputDTMs generated using different methods, the merits of these algorithms can be retained whilst theirlimitations can be offset effectively. In future studies, researchers should continue to work on moreapproaches for combining different ground filtering methods. In addition, it is of practical significanceto examine which methods combined together may achieve the optimum results.4.2.2. Advanced DTM Generation Using Multi-SourcesAlthough great progress has been made in the design of new DTM generation methods and somechallenging issues have been addressed to some extent, some limitations remain. Specifically, it isdifficult for traditional methods to precisely generate DTMs in complicated terrain situations thatare characterized with sharp terrain relief and a diversity of non-ground features. Due to its rapidlychanging elevation, sharp terrain relief is very likely to be considered as non-ground objects accordingto morphological filters, slope operators or statistical functions. Bringing in additional features toassist the process of ground point filtering is a practical solution to this problem. Although somescholars conduct image segmentation and classification and aim to classify non-ground objects usingthe intensity feature, the single band of intensity attribute cannot support the task of classifying adiversity of non-ground features. Moreover, as explained, the intensity feature provided by Lidardata is not fully reliable for classification. Furthermore, the existence of discontinuity and highlyfragmented landscapes inevitably prevents a satisfactory DTM generation result using airborne Lidardata solely, especially in large study sites with complicated terrain and object conditions.In recent years, the integration of Lidar data and other data sources for land use (cover)classification has gained gpopularity. In addition, it is widely accepted that integrating Lidar datawith multi-spectral remote sensing data achieves much better classification accuracy than the soleuse of airborne Lidar. As a result, Lidar data has been fused with multi-spectral images [92–94],high-resolution imagery [95–98], airborne photography [99] and hyperspectral imagery [100–102] forSensors 2017, 17, 150 18 of 24land cover classification. Amongst these studies, most scholars considered the elevation feature fromLidar data for better classification accuracy, yet few studies employed additional data sources for betterDTM generation. This imbalance may result from people’s biased understanding of DTM generationand airborne Lidar data. Many scholars believe that the main application of airborne Lidar data isDTM generation, and the sole use of airborne Lidar is more than enough for DTM generation. In fact,including additional spectral features provided by other data sources is probably a better solutionfor DTM generation in complicated terrains than designing advanced DTM methods using Lidardata only. Recently, an exciting development that some scholars started to apply is multiple sourcesfor better DTM generation. Kim et al. employed the spectral feature from aerial images [103] andhigh-resolution satellite images [104] to better classify bare ground and building roofs, which is onecommon challenge in urban DTM generation. Nordbo et al. [105] employed existing building mask(geographical information on the coastline and building edges) to assist the generation of large-scaleurban DTMs using airborne Lidar data. Debella-Gilo [106] combined high-resolution photogrammetricpoint clouds (similar to Lidar point clouds) with existing low-resolution DTMs, which worked asreference trend surface, to produce high-resolution DTMs. Instead of direct use of spectral information,Saeidi et al. [107] extracted the NDVI (Normalized Difference Vegetation Index) feature to assistthe classification of bare ground, trees attached to steep terrains and large buildings in complicatedlandscapes and produced high-quality DTMs. Since the Laser pulses can be absorbed when hitting thewater surface, geomorphic analysis at coastal regions using Lidar data solely is difficult. Therefore, theuse of both airborne Lidar and imaging sensors is highly recommended [91,108]. Moretto et al. [109]fused Lidar, colour bathymetry and dgps surveys to better feature coastal line terrain, which is a typicaldemonstration of employing multiple sources to generate DTMs in complicated terrains. However,the implementation of multi-source supported DTM generation should be further explored throughmany more case studies conducted in complicated and highly fragmented landscapes with steep anddiscontinuous terrains and a variety of attached objects.Full-waveform Lidar data has been increasingly examined, as it provides additional spatialcorrelation and useful features. Mallet et al. [9,10] classified ground, vegetation and building pointsusing such specific features as differential laser cross-section from full-waveform Lidar data, whilstJutzi and Stilla [8] conducted urban land cover classification using the neighborhood relationshipsbetween waveforms. Considering that full-waveform Lidar data can provide many additionalfeatures, this data source is of great potential for DTM generation in very complicated or highlyfragmented terrain situations, which requires a diversity of features to distinguish steep terrain fromnon-ground objects.5. SummaryOver the past few decades, methods for digital terrain model (DTM) generation have beenintensively studied and many algorithms have been proposed to derive DTMs under differentterrain situations. This paper reviews recent progress of DTM generation and categorized existingDTM generation methods into six major classes: surface-based adjustment, morphology-basedfiltering, TIN-based refinement, segmentation and classification, statistical analysis and multi-scalecomparison. The principle, typical algorithms, suitability and limitations for each category areexplained and compared.Since DTM generation methods of different categories have their own strengths and limitationsin specific terrains, the simple use of one type of ground point filter can hardly satisfy requirementsof a diversity of terrains. To this end, some new ground point filters have been proposed recently tomake generally applicable DTM generation possible. These algorithms have successfully combinedseveral categories of ground point filters, and can thus retain terrain details and remove a diversity ofobjects. Therefore, these new methods proved their capability of producing reliable DTMs even undermany challenging terrains, indicating a significant progress in the principle and implementations ofDTM generation.Sensors 2017, 17, 150 19 of 24Despite the development in the reliability and generalization of DTM generation, some limitationsstill exist. Due to limited features provided by light detection and ranging (Lidar) data, it remainsvery difficult, if not impossible, to simultaneously distinguish complicated terrain situations (e.g.,discontinuities and shape ridges), highly fragmented landscapes, and a variety of objects using airborneLidar data solely. This challenge is inevitable in DTM generation implementation conducted in largeareas, which include a variety of complicated terrain relief and non-ground objects. An effectivesolution is to integrate additional sources with airborne Lidar data for better DTM generation.Although the fusion of multi-spectral images with Lidar data has been frequently implementedin land cover/use classification and other fields, most scholars only considered the elevation featurefrom Lidar data for better classification whilst few studies employed additional sources for betterDTM generation. Several recent studies have started to combine Lidar data with additional sourcesfor generating DTMs in complicated terrains, which should be a promising direction for furtherdevelopment. The selection of data sources, the method for data fusion, and the suitability of terrainsare the key issues for multi-source supported DTM generation, which should be explored throughmany more case studies.This review presented recent developments, remaining challenges and promising directions forDTM generation using airborne Lidar data. It also provides useful references for scholars to properlychoose methods according to specific terrains, or design new methods based on a better understandingof existing algorithms.Acknowledgments: This research is supported by National Natural Science Foundation of China (Grant No.210100066), the National Key Research and Development Program of China (No. 2016YFA0600104) and BeijingTraining Support Project for Excellent Scholars 2015000020124G059.Author Contributions: Ziyue Chen and Bernard Devereux contributed to the design and writing. Bingbo Gaocontributed to the writing.Conflicts of Interest: The authors declare no conflict of interest.References1. Sithole, G.; Vosselman, G. Comparison of filtering algorithms. Int. Arch. Photogramm. Remote Sens. Spat.Inf. Sci. 2003, 34, 71–78.2. Chen, Z.Y.; Devereux, B.; Gao, B.B.; Amable, G. Upward-fusion urban DTM generating method usingairborne Lidar data. ISPRS J. Photogramm. Remote Sens. 2012, 72, 121–130. [CrossRef]3. Lu, W.; Murphy, K.P.; Little, J.J.; Sheffer, A.; Fu, H. A Hybrid Conditional Random Field for Estimatingthe Underlying Ground Surface from Airborne LiDAR Data. IEEE Trans. Geosci. Remote Sens. 2009, 47,2913–2922.4. Vu, T.T.; Tokunaga, M.; Yamazaki, F. LiDAR signatures to update Japanese building inventory database.In Proceedings of the 25th Asian Confence on Remote Sensing, Chiang Mai, Thailand, 22–26 November 2004.5. Zhang, Y.; Men, L. Study of the airborne LIDAR data filtering methods. In Proceedings of the InternationalConference on Geoinformatics: Giscience in Change, Beijing, China, 18–20 June 2010.6. Liu, X.Y. Airborne LiDAR for DEM generation: Some critical issues. Prog. Phys. Geogr. 2008, 32, 31–49.7. Meng, X.; Currit, N.; Zhao, K. Ground Filtering Algorithms for Airborne LiDAR Data: A Review of CriticalIssues. Remote Sens. 2010, 2, 833–860. [CrossRef]8. Jutzi, B.; Stilla, U. Waveform processing of laser pulses for reconstruction of surfaces in urban areas.In Proceedings of the Urban 2005, Tempe, AZ, USA, 14–16 March 2005.9. Mallet, C.; Bretar, F.; Soergel, U. Analysis of full-waveform Lidar data for classification of urban areas.Photogramm. Fernerkund. Geoinform. 2008, 5, 337–349.10. Mallet, C.; Bretar, F.; Roux, M.; Soergel, U.; Heipke, C. Relevance assessment of full-waveform lidar data forurban area classification. ISPRS J. Photogramm. Remote Sens. 2011, 66, S71–S84. [CrossRef]11. Doneus, M.; Briese, C. Full-waveform airborne laser scanning as a tool for archaeological reconnaissance.BAR Int. Ser. 2006, 1568, 99–106.Sensors 2017, 17, 150 20 of 2412. Doneus, M.; Briese, C. Digital terrain modelling for archaeological interpretation within forested areasusing fullwaveform laser scanning. In Proceedings of the 7th International Symposium on Virtual Reality,Archaeology and Cultural Heritage VAST, Nicosia, Cyprus, 30 October–4 November 2006.13. Meng, X.; Wang, L.; Silvan-Cardenas, J.L.; Currit, N. A multidirectional ground filtering algorithm forairborne Lidar. ISPRS J. Photogramm. Remote Sens. 2009, 64, 117–124. [CrossRef]14. Silvan-Cardenas, J.L.; Wang, L. A multi-resolution approach for filtering Lidar altimetry data. ISPRS J.Photogramm. Remote Sens. 2006, 61, 11–22. [CrossRef]15. Meng, X.; Wang, L.; Currit, N. Morphology-based building detection from airborne LIDAR data.Photogramm. Eng. Remote Sens. 2009, 75, 437–442. [CrossRef]16. Chen, Q.; Gong, P.; Baldocchi, D.; Xie, G. Filtering airborne laser scanning data with morphological methods.Photogramm. Eng. Remote Sens. 2007, 73, 175–185. [CrossRef]17. Kobler, A.; Pfeifer, N.; Ogrinc, P.; Todorovski, L.; Ostir, K.; Dzeroski, S. Repetitive interpolation: A robustalgorithm for DTM generation from aerial laser scanner data in forested terrain. Remote Sens. Environ. 2007,108, 9–23. [CrossRef]18. Breunig, M.M.; Kriegel, H.P.; Ng, R.T.; Sander, J. LOF: Identifying density-based local outliers. In Proceedingsof the ACM SIGMOD International Conference on Management of Data, Dallas, TX, USA, 15–18 May 2000;pp. 93–104.19. Sotoodeh, S. Hierarchical clustered outlier detection in laser scanner point clouds. Int. Arch. Photogramm.Remote Sens. 2007, 35, 383–388.20. Vain, A.; Kaasalainen, S. Correcting airborne laser scanning intensity data. In Laser Scanning, Theory andApplications; Wang, C.-C., Ed.; InTech: Rijeka, Croatia, 2011.21. Goovaerts, P. Geostatistics for Natural Resources Evaluation; Oxford University Press: New York, NY, USA,1997; pp. 59–63.22. Isaaks, E.H.; Srivastava, R.M. An Introduction to Applied Geostatistics; Oxford University Press: New York, NY,USA, 1989.23. Fisher, P.F.; Tate, N.J. Causes and consequences of error in digital elevation models. Prog. Phys. Geogr. 2006,30, 467–489. [CrossRef]24. Kraus, K.; Pfeifer, N. Determination of terrain models in wooded areas with airborne laser scanner data.ISPRS J. Photogramm. Remote Sens. 1998, 53, 193–203. [CrossRef]25. Kraus, K.; Pfeifer, N. Advanced DTM generation from LIDAR data. Int. Arch. Photogramm. Remote Sens.2001, 34, 23–30.26. Elmqvist, M. Ground surface estimation from airborne laser scanner data using active shape models.Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2002, 34, 114–118.27. Pfeifer, N.; Stadler, P.; Briese, C. Derivation of Digital Terrain Models in the SCOP ++ Environment.In Proceedings of the OEEPE Workshop on Airborne Laser Scanning and Interferometric SAR for DetailedDigital Elevation Models, Stockholm, Sweden, 1–3 March 2001.28. Wack, R.; Wimmer, A. Digital terrain models from airborne laser scanner data—A grid based approach.Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2002, 34, 293–296.29. Chen, H.Y.; Cheng, M.S.; Li, J.; Liu, Y. An iterative terrain recovery approach to automated DTM. Int. Arch.Photogramm. Remote Sens. Spat. Inf. Sci. 2012, 39, 363–368. [CrossRef]30. Maguya, A.S.; Junttila, V.; Kauranne, T. Adaptive algorithm for large scale dtm interpolation from lidardata for forestry applications in steep forested terrain. ISPRS J. Photogramm. Remote Sens. 2013, 85, 74–83.[CrossRef]31. Zhang, W.M.; Qi, J.B.; Wan, P.; Wang, H.T.; Xie, D.H.; Wang, X.Y.; Yan, G.J. An Easy-to-Use Airborne LiDARData Filtering Method Based on Cloth Simulation. Remote Sens. 2016, 8, 501. [CrossRef]32. Mongus, D.; Zalik, B. Computationally Efficient Method for the Generation of a Digital Terrain Model fromAirborne LiDAR Data Using Connected Operators. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2014, 7,340–351. [CrossRef]33. Vosselman, G. Slope based filtering of laser altimetry data. Int. Arch. Photogramm. Remote Sens. 2000, 3,935–942.34. Sithole, G. Filtering of laser altimetry data using a slope adaptive filter. Int. Arch. Photogramm. Remote Sens.Spat. Inf. Sci. 2001, 34, 203–210.Sensors 2017, 17, 150 21 of 2435. Roggero, M. Dense DTM from laser scanner data. In Proceedings of the OEPEE Workshop on Airborne LaserScanning and Interferometric SAR, Stockholm, Sweden, 1–3 March 2001.36. Zakšek, K.; Pfeifer, N. An Improved Morphological Filter for Selecting Relief Points from a LIDAR Point Cloudin Steep Areas with Dense Vegetation; Technical Report; Institute of Anthropological and Spatial Studies:Ljubljana, Solvenia, 2006.37. Shao, Y.; Chen, L. Automated Searching of Ground Points from Airborne Lidar Data Using a Climbing andSliding Method. Photogramm. Eng. Remote Sens. 2008, 74, 625–635. [CrossRef]38. Lohmann, P.; Koch, A.; Schaeffer, M. Approaches to the filtering of laser scanner data. Int. Arch. Photogramm.Remote Sens. 2000, 33, 540–547.39. Shan, J.; Sampath, A. Urban DEM Generation from Raw Lidar Data: A Labeling Algorithm and ItsPerformance. Photogramm. Eng. Remote Sens. 2005, 71, 217–226. [CrossRef]40. Wang, C.; Tseng, Y. Dual-directional profile filter for digital terrain model generation from airborne laserscanning data. J. Appl. Remote Sens. 2014, 8, 4480–4494. [CrossRef]41. Hu, X.; Ye, L.; Pang, S.; Shan, J. Semi-global filtering of airborne lidar data for fast extraction of digital terrainmodels. Remote Sens. 2015, 7, 10996–11015. [CrossRef]42. Li, Y.; Yong, B.; Wu, H.; An, R.; Xu, H. An improved top-hat filter with sloped brim for extracting groundpoints from airborne lidar point clouds. Remote Sens. 2014, 6, 12885–12908. [CrossRef]43. Susaki, J. Adaptive Slope Filtering of Airborne LiDAR Data in Urban Areas for Digital Terrain Model (DTM)Generation. Remote Sens. 2012, 4, 1804–1819. [CrossRef]44. Pingel, T.J.; Clarke, K.C.; McBride, W.A. An improved simple morphological filter for the terrain classificationof airborne LIDAR data. ISPRS J. Photogramm. Remote Sens. 2013, 77, 21–30. [CrossRef]45. Axelsson, P. DEM generation form laser scanner data using adaptive TIN models. Int. Arch. Photogramm.Remote Sens. 2000, 4, 110–118.46. Sohn, G.; Dowman, I. Terrain surface reconstruction by the use of tetrahedron model with the mdl criterion.Int. Arch. Photogramm. Remote Sens. 2002, 34, 336–344.47. Guan, H.Y.; Li, J.; Yu, Y.T.; Zhong, L.; Ji, Z. DEM generation from lidar data in wooded mountain areas bycross-section-plane analysis. Int. J. Remote Sens. 2014, 35, 927–948. [CrossRef]48. Chen, Q.; Wang, H.; Zhang, H.; Sun, M.; Liu, X. A point cloud filtering approach to generating DTMs forsteep mountainous areas and adjacent residential areas. Remote Sens. 2016, 8, 71. [CrossRef]49. Zhang, J.; Lin, X. Filtering airborne LiDAR data by embedding smoothness-constrained segmentation inprogressive TIN densification. ISPRS J. Photogramm. Remote Sens. 2013, 81, 44–59. [CrossRef]50. Brovelli, M.A.; Cannata, M.; Longoni, U. Lidar data filtering and dtm interpolation within grass. Trans. GIS2004, 8, 155–174. [CrossRef]51. Filin, S.; Pfeifer, N. Segmentation of Airborne Laser Scanning Data Using a Slope Adaptive Neighbourhood.ISPRS J. Photogramm. Remote Sens. 2006, 60, 71–80. [CrossRef]52. Lohmann, P. Segmentation and filtering of laser scanner digital surface models. Int. Arch. Photogramm.Remote Sens. Spat. Inf. Sci. 2002, 34, 311–316.53. Nardinocchi, C.; Forlani, G.; Zingaretti, P. Classification and filtering of laser data. In Proceedings of theISPRS working group III/3 workshop 3-D Reconstruction from Airborne Laser Scanner and InSAR Data,Dresden, Germany, 8–10 October 2003.54. Niemeyer, J.; Rottensteiner, F.; Soergel, U. Contextual classification of lidar data and building object detectionin urban areas. ISPRS J. Photogramm. Remote Sens. 2014, 87, 152–165. [CrossRef]55. Roggero, M. Airborne laser scanning: Clustering in raw data. Int. Arch. Photogramm., Remote Sens. Spat.Inf. Sci. 2001, 34, 227–232.56. Roggero, M. Object segmentation with region growing and principal component analysis. Int. Arch.Photogramm. Remote Sens. Spat. Inf. Sci. 2002, 34, 289–294.57. Sithole, G.; Vosselman, G. Filtering of airborne laser scanner data based on segmented point clouds.Workshop Laser Scanning 2005, 36, W19.58. Tovari, D.; Pfeifer, N. Segmentation based robust interpolation-a new approach to laser data filtering.Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2005, 36, 79–84.59. Brennan, R.; Webster, T.L. Object-oriented land cover classification of lidar derived surfaces. Can. J.Remote Sens. 2006, 32, 162–172. [CrossRef]Sensors 2017, 17, 150 22 of 2460. Buján, S.; González-Ferreiro, E.; Reyes-Bueno, F.; Barreiro-Fernández, L.; Crecente, R.; Miranda, D. Landuse classification from LiDAR data and ortho-images in a rural area. Photogramm. Rec. 2012, 27, 401–422.[CrossRef]61. Chen, Z.Y.; Xu, B.; Gao, B.B. An Image-Segmentation-Based Urban DTM Generation Method Using AirborneLidar Data. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2016, 9, 496–506. [CrossRef]62. Kato, A.; Moskal, L.M.; Schiess, P.; Swanson, M.E.; Calhoun, D.; Stuetzle, W. Capturing tree crown formationthrough implicit surface reconstruction using airborne lidar data. Remote Sens. Environ. 2009, 113, 1148–1162.[CrossRef]63. Lee, A.C.; Lucas, R.M. A LiDAR-derived canopy density model for tree stem and crown mapping inAustralian forests. Remote Sens. Environ. 2007, 111, 493–518. [CrossRef]64. Blaschke, T. Object based image analysis for remote sensing. ISPRS J. Photogramm. Remote Sens. 2010, 65,2–16. [CrossRef]65. Blaschke, T.; Tomljenovic, I. LIDARScapes and OBIA. In Proceedings of the ASPRS Annual Conference,Sacramento, CA, USA, 19–23 March 2012.66. Antonarakis, A.S.; Richards, K.S.; Brasington, J. Object-based land cover classification using airborne LiDAR.Remote Sens. Environ. 2008, 112, 2988–2998. [CrossRef]67. Johansen, K.; Arroyo, L.A.; Armston, J.; Phinn, S.; Witte, C. Mapping riparian condition indicators ina sub-tropical savanna environment from discrete return lidar data using object-based image analysis.Ecol. Indic. 2010, 10, 796–807. [CrossRef]68. Im, J.; Jensen, J.R.; Hodgson, M.E. Object-based land cover classification using high-posting-density LiDARdata. GIScience Remote Sens. 2008, 45, 209–228. [CrossRef]69. Samadzadegan, F.; Bigdeli, B.; Ramzi, P. A multiple classifier system for classification of LiDAR remotesensing data using multi-class SVM. In Proceedings of the 9th International Conference on Multiple Classifiersystems, Cairo, Egypt, 7–9 April 2010; pp. 254–263.70. Huang, X.; Zhang, L.; Gong, W. Information fusion of aerial images and LiDAR data in urban areas:Vector-stacking, re-classification and post-processing approaches. Int. J. Remote Sens. 2011, 32, 69–84.[CrossRef]71. Chen, Z.Y.; Gao, B.B. An Object-Based Method for Urban Land Cover Classification Using Airborne LidarData. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2014, 10, 4243–4354. [CrossRef]72. Zhang, J.; Lin, X.; Ning, X. SVM-Based Classification of Segmented Airborne LiDAR Point Clouds in UrbanAreas. Remote Sens. 2013, 5, 3749–3775. [CrossRef]73. Bartels, M.; Wei, H. Towards DTM generation from LIDAR data in hilly terrain using wavelets. In Proceedingsof the 4th International Workshop on Pattern Recognition in Remote Sensing in Conjunction with ICPR 2006,Hong Kong, China, 20–24 August 2006; pp. 33–36.74. Bretar, F.; Chehata, N. Terrain Modeling from Lidar Range Data in Natural Landscapes: A Predictive andBayesian Framework. IEEE Trans. Geosci. Remote Sens. 2010, 48, 1568–1578. [CrossRef]75. Bartels, M.; Wei, H. Threshold-free object and ground point separation in LIDAR data. Pattern Recognit. Lett.2010, 31, 1089–1099. [CrossRef]76. Yao, W.; Hinz, S.; Stilla, U. Automatic vehicle extraction from airborne LiDAR data of urban areas usingmorphological reconstruction. In Proceedings of the 5th IAPRS Workshop on Pattern Recognition in RemoteSensing (PRRS 2008), Tampa, FL, USA, 7 December 2008; p. 4.77. Bao, Y.F.; Li, G.P.; Cao, C.X.; Li, X.W.; Zhang, H.; He, Q.S.; Bai, L.Y.; Chang, C.Y. Classification of LIDAR pointcloud and generation of DTM from LIDAR height and intensity data in forested area. Int. Arch. Photogramm.Remote Sens. Spat. Inf. Sci. 2008, 37, 313–318.78. Mongus, D.; Zalik, B. Parameter-free ground filtering of LiDAR data for automatic DTM generation. ISPRS J.Photogramm. Remote Sens. 2012, 67, 1–12. [CrossRef]79. Zhang, K.Q.; Chen, S.C.; Whitman, D.; Shyu, M.L.; Yan, J.H.; Zhang, C.C. A progressive morphological filterfor removing nonground measurements from airborne LIDAR data. IEEE Trans. Geosci. Remote Sens. 2003,41, 872–882. [CrossRef]80. Li, S.W.; Sun, H.B.; Yan, L. A Filtering Method for Generating DTM based on Multi- scale MathematicMorphology. In Proceedings of the 2011 IEEE. International Conference on Mechatronics and Automation,Beijing, China, 7–10 August 2011; pp. 693–697.Sensors 2017, 17, 150 23 of 2481. Xiong, J.H.; Fang, Y.M.; Jin, B.X.; Zhao, Z.F. Automated DTM Generation in Urban Areas with AirborneLiDAR Data. In Proceedings of the 4th International Conference on Intelligent Human-Machine Systemsand Cybernetics, Nanchang, China, 26–27 August 2012; pp. 181–184.82. Chen, C.F.; Li, Y.Y.; Li, W.; Dai, H.L. A multiresolution hierarchical classification algorithm for filteringairborne LiDAR data. ISPRS J. Photogramm. Remote Sens. 2013, 82, 1–9. [CrossRef]83. Mongus, D.; Lukaˇc, N.; Žalik, B. Ground and building extraction from LiDAR data based on differentialmorphological profiles and locally fitted surfaces. ISPRS J. Photogramm. Remote Sens. 2014, 7, 145–156.[CrossRef]84. Maguya, A.S.; Junttila, V.; Kauranne, T. Algorithm for Extracting Digital Terrain Models under Forest Canopyfrom Airborne LiDAR Data. Remote Sens. 2014, 6, 6524–6548. [CrossRef]85. Su, W.; Sun, Z.; Zhong, R.; Huang, J.; Li, M.; Zhu, J.; Zhang, K.; Wu, H.; Zhu, D. A new hierarchical movingcurve-fitting algorithm for filtering lidar data for automatic DTM generation. Int. J. Remote Sens. 2015, 36,3616–3635. [CrossRef]86. Mandlburger, G.; Pfeifer, N.; Ressl, C.; Briese, C.; Roncat, A.; Lehner, H.; Mucke, W. Algorithms and Toolsfor Airborne LiDAR Data Processing from a Scientific Perspective; ELMF World Forum 2010: The Hague,The Netherlands, 2010.87. Forczmanski, P.; Maleika, W. Predicting the Number of DCT Coefficients in the Process of Seabed DataCompression. In Lecture Notes in Computer Science, Proceeding of the Computer Analysis of Images and Patterns,CAIP 2015, Ystad, Sweden, 2–4 September 2015; Springer International Publishing: Cham, Switzerland;Volume 9256, pp. 77–87.88. Forczmanski, P.; Maleika, W. Near-Lossless PCA-Based Compression of Seabed Surface with Prediction.In Lecture Notes in Computer Science, Proceeding of the Image Analysis and Recognition (ICIAR 2015), NiagaraFalls, ON, Canada, 22–24 July 2015; Springer International Publishing: Cham, Switzerland; Volume 9164,pp. 119–128.89. Quintero, R.; Guzman, G.; Torres, M.; Menchaca-Mendez, R.; Moreno-Ibarra, M.; Mata, F. A CompressionAlgorithm for Managing Digital Elevation Models in Mobile Devices. J. Univers. Comput. Sci. 2014, 20,1433–1442.90. Scarmana, G. Lossless data compression of grid-based digital elevation models: A PNG image formatevaluation. In Proceedings of the ISPRS Technical Commission V Symposium, Close-Range Imaging,Ranging and Applications, Riva del Garda, Italy, 23–25 June 2014.91. Mohammadzadeh, A.; Valadan Zoej, M.J. A State of Art on Airborne Lidar Application in Hydrology andOceanography: A Comprehensive Overview. Int. Arch. Photogramm. Remote Sens. Spat. Inf. Sci. 2008, 39,315–320.92. Rottensteiner, F.; Trinder, J.; Clode, S.; Kubik, K. Using the Dempster–Shafer method for the fusion of LIDARdata and multi-spectral images for building detection. Inf. Fusion 2005, 6, 283–300. [CrossRef]93. Bork, E.; Su, J. Integrating LIDAR data and multispectral imagery for enhanced classification of rangelandvegetation: A meta analysis. Remote Sens. Environ. 2007, 111, 11–24. [CrossRef]94. Singh, K.; Vogler, J.B.; Shoemaker, D.A.; Meentemeyer, R.K. LiDAR-Landsat data fusion for large-areaassessment of urban land cover: Balancing spatial resolution, data volume and mapping accuracy. ISPRS J.Photogramm. Remote Sens. 2012, 74, 110–121. [CrossRef]95. Chen, Y.H.; Su, W.; Li, J.; Sun, Z.P. Hierarchical object oriented classification using very high resolutionimagery and LIDAR data over urban areas. Adv. Space Res. 2009, 43, 1101–1110. [CrossRef]96. Ke, Y.H.; Quackenbush, L.J.; Im, J. Synergistic use of QuickBird multispectral imagery and LIDAR data forobject-based forest species classification. Remote Sens. Environ. 2010, 114, 1141–1154. [CrossRef]97. Syed, S.; Dare, P.; Jones, S. Automatic Classification of Land Cover Features with High Resolution Imageryand LIDAR Data: An Object-Oriented Approach. Available online: http://www.ecognition.com/sites/default/files/266_0185.pdf.2005 (accessed on 1 September 2005).98. Teo, T.A.; Chen, L.C. Object-based building detection from LiDAR data and high resolution satellite imagery.In Proceedings of the Asian Conference on Remote Sensing, Ching-Mai, Thailand, 22–26 November 2004.99. Hill, R.A.; Thomson, A.G. Mapping woodland species composition and structure using airborne spectraland LiDAR data. Int. J. Remote Sens. 2005, 26, 3763–3779. [CrossRef]100. Dalponte, M.; Bruzzone, L.; Gianelle, D. Fusion of Hyperspectral and LIDAR Remote Sensing Data forClassification of Complex Forest Areas. IEEE Trans. Geosci. Remote Sens. 2008, 46, 416–1427. [CrossRef]Sensors 2017, 17, 150 24 of 24101. Dalponte, M.; Bruzzone, L.; Gianelle, D. Tree species classification in the Southern Alps based on the fusion ofvery high geometrical resolution multispectral/hyperspectral images and LiDAR data. Remote Sens. Environ.2012, 123, 258–270. [CrossRef]102. Luo, S.; Wang, C.; Xi, X.; Zeng, H.; Li, D.; Xia, S.; Wang, P. Fusion of Airborne Discrete-Return LiDAR andHyperspectral Data for Land Cover Classification. Remote Sens. 2016, 8, 3. [CrossRef]103. Kim, Y.M.; Eo, Y.D.; Chang, A.J.; Kim, Y.I. Generation of a DTM and building detection based on an MPFthrough integrating airborne lidar data and aerial images. Int. J. Remote Sens. 2013, 34, 2947–2968. [CrossRef]104. Kim, Y.M.; Kim, Y. Improved Classification Accuracy Based on the Output-Level Fusion of High-ResolutionSatellite Images and Airborne LiDAR Data in Urban Area. IEEE Geosci. Remote Sens. Lett. 2014, 11, 636–640.105. Nordbo, A.; Karsisto, P.; Matikainen, L.; Wood, C.R.; Järvi, L. Urban surface cover determined with airbornelidar at 2 m resolution—Implications for surface energy balance modelling. Urban Clim. 2015, 13, 52–72.[CrossRef]106. Debella-Gilo, M. Bare-earth extraction and DTM generation from photogrammetric point clouds includingthe use of an existing lower-resolution DTM. Int. J. Remote Sens. 2016, 37, 3104–3124. [CrossRef]107. Saeidi, V.; Pradhan, B.; Idrees, M.O.; Abd Latif, Z. Fusion of airborne lidar with multispectral spot 5 imagefor enhancement of feature extraction using dempster–shafer theory. IEEE Trans. Geosci. Remote Sens. 2014,52, 6017–6025. [CrossRef]108. Toth, C.K. The State-of-the-Art in Airborne Data Collection Systems—Focused on LiDAR and Optical Imager.In Proceedings of the 100th Anniversary Photogrammetric Week (PhoWo 09), Stuttgart, Germany, 7–11September 2009.109. Moretto, J.; Rigon, E.; Mao, L.; Delai, F.; Picco, L.; Lenzi, M.A. Short-term geomorphic analysis in a disturbedfluvial environment by fusion of lidar, colour bathymetry and dgps surveys. Catena 2014, 122, 180–195.[CrossRef]© 2017 by the authors; licensee MDPI, Basel, Switzerland. This article is an open accessarticle distributed under the terms and conditions of the Creative Commons Attribution(CC-BY) license (http://creativecommons.org/licenses/by/4.0/).

Let’s block ads! (Why?)

Do you need any assistance with this question?
Send us your paper details now
We’ll find the best professional writer for you!

 



error: Content is protected !!