Cloth simulation-based construction of pit-free canopy height models from airborne LiDAR data
- 222 Downloads
The universal occurrence of randomly distributed dark holes (i.e., data pits appearing within the tree crown) in LiDAR-derived canopy height models (CHMs) negatively affects the accuracy of extracted forest inventory parameters.
We develop an algorithm based on cloth simulation for constructing a pit-free CHM.
The proposed algorithm effectively fills data pits of various sizes whilst preserving canopy details. Our pit-free CHMs derived from point clouds at different proportions of data pits are remarkably better than those constructed using other algorithms, as evidenced by the lowest average root mean square error (0.4981 m) between the reference CHMs and the constructed pit-free CHMs. Moreover, our pit-free CHMs show the best performance overall in terms of maximum tree height estimation (average bias = 0.9674 m).
The proposed algorithm can be adopted when working with different quality LiDAR data and shows high potential in forestry applications.
KeywordsData pits Tree crown Canopy height models Cloth simulation Pit-free
Canopy Height Model
Digital Surface Model
Digital Terrain Model
Global Positioning System
Airborne Light Detection and Ranging
Root Mean Square Error
In the last two decades, airborne light detection and ranging (LiDAR) has become a reliable remote sensing technique for forest inventory given its capability to provide precise and detailed three-dimensional (3D) information on forest structures directly (Lim et al. 2003; Popescu 2007; Hyyppä et al. 2008; Yu et al. 2011; Huang and Lian 2015; Coomes et al. 2018; Stereńczak et al. 2018). The construction of a canopy height model (CHM) from LiDAR data is an effective means of increasing data usability (Mielcarek et al. 2018). The CHM represents an absolute canopy height above the terrain surface and is usually constructed by subtracting the digital terrain model (DTM) from the digital surface model (DSM) (Khosravipour et al. 2014). The CHM quality significantly affects the estimation accuracy of forest inventories and subsequent biophysical parameters (Shamsoddini et al. 2013).
A common problem in the LiDAR-derived CHM is the occurrence of unnatural black holes (i.e., data pits appearing within the tree crown) (Mielcarek et al. 2018). Data pits exhibit irregular height variations, and their heights are lower than those of their neighbours in raster CHM (Khosravipour et al. 2014; Mielcarek et al. 2018). In a CHM with a certain resolution (< 2 m), data pits are usually single or several pixel(s) clumping together, that is, the data pit size is irregular (Zhao et al. 2013). Existing literature has indicated no specific causes of data pits. However, many researchers have inferred various possible causes. Data pits exist due to the combination of multiple factors, from data collection to post-processing. Shamsoddini et al. (2013) reported that the positioning error of the Global Positioning System (GPS) is a key cause of data pits. A significant GPS vertical error causes data pits when flight heights are < 3000 m above ground and scan angles are outside 9° (Latypov 2005; Goulden and Hopkinson 2010). Leckie et al. (2003) found that data pits exist because of the integration of different flight line datasets, the penetration of laser beams through canopy branches and foliage into the ground and the unevenly distributed spatial resolution of LiDAR point clouds caused by a scanning mechanism, variation in aircraft attitude and deflection of lost returns. Ben-Arie et al. (2009) inferred that, when the points are binned to a grid, many grid cells will remain empty because some information from points with similar x–y coordinates but different z values are lost. After being interpolated to a raster, pits may be generated in empty cells (Axelsson 1999). Khosravipour et al. (2014) concluded that data pits may be formed during ground filtering when generating a DSM or a DTM, depending on filtering technique and LiDAR point density (Kraus and Pfeifer 1998). Previous studies have asserted that such pits have a seriously negative effect on forest parameter estimation because these pits disrupt the CHMs. Gaveau and Hill (2003) indicated that data pits cause the underestimation of tree height for LiDAR data with a mean point density of 0.21 points ∙ m–2. Shamsoddini et al. (2013) found that data pits negatively affect the estimation of basal area and stand volume for LiDAR data with a mean point density of 2 pulses ∙ m–2. Khosravipour et al. (2014) demonstrated that data pits lead to large omission errors (undetected trees) and commission errors (falsely detected trees) in treetop detection for LiDAR data with mean point densities of 7 and 160 points ∙ m–2. Thus, data pits must be filled to ensure that the uppermost layer of the forest canopy is realistically represented by the LiDAR-derived CHM.
Many algorithms have been proposed to construct pit-free CHMs. These algorithms can be divided into two categories, namely, raster- and point cloud-based methods. Raster-based methods are achieved by interpolating point clouds into raster images. The data pits are then filled by the classical image processing methods, such as mean and median filters. These filters are simple and fast, but the result is sensitive to the optimal kernel size given the pit size irregularity (Shamsoddini et al. 2013; Mielcarek et al. 2018).
Kernel size selection is critical to the success of many image processing methods. For example, when using morphological-based ground filtering algorithm, only small non-ground objects, such as trees, will be effectively removed by small kernels; however, this algorithm tends to over-remove the ground points when using large kernels (Zhang et al. 2003). Similarly, for mean and median filters, only data pits with a small size will be effectively filled by small kernels; however, these filters tend to over-smooth the CHM when using large kernels (Ben-Arie et al. 2009).
These filters negatively affect the accuracy of measurements, such as underestimation of tree heights and crown radius, deriving from the CHM. Such a phenomenon occurs because the elevation values of other pixels in CHM, except for the pitted pixels, are also changed, leading to treetop omission and crown shoulder reduction (Khosravipour et al. 2014). Accordingly, several solutions have been proposed to fill data pits without affecting other CHM values. Shamsoddini et al. (2013) detected data pits by computing a similarity index in a 3 × 3 kernel and then used a mean filter with a 5 × 5 kernel to fill such pits. Ben-Arie et al. (2009) used a Laplacian operator with a 3 × 3 kernel to detect data pits and then utilised a median filter with a 3 × 3 kernel to fill such pits. These improved algorithms only fill the pitted pixels without affecting other pixel values in CHM. However, they cannot solve the kernel size selection problem. Therefore, these algorithms may be ineffective for large pits because they use a fixed kernel size.
Another type of pit-free CHM construction methods (i.e., point cloud-based methods) directly works on height-normalised point clouds rather than raw CHMs. Khosravipour et al. (2014) constructed a stack of partial CHMs from different height intervals of the canopy; subsequently, these CHMs were combined into one CHM on the basis of the highest value across all CHMs for each x and y raster position. However, this algorithm must carefully determine a series of complicated parameters for filling pits at different heights. Chen et al. (2017) used locally weighted regression and z-scores to construct a pit-free CHM. This algorithm is highly automatic but may not work for pits in sparse density data because this algorithm requires calculation in the neighbourhood. The results were related to neighbourhood size.
The use of the aforementioned algorithms has proven to be successful. However, the prominent limitations of existing algorithms are presented as follows: (a) having applicability only to LiDAR data with regular-sized pits, (b) changing the values of not only the pitted pixels but also other CHM ones or (c) requiring numerous complicated parameters. To cope with these problems, we propose a novel algorithm based on cloth simulation for constructing a pit-free CHM. The specific objectives of this study are to (a) evaluate the applicability of the proposed algorithm to different proportions of data pits, (b) assess cloth simulation-based CHMs through visual comparison of the mean- and median-filtered CHMs, and (c) evaluate and compare the accuracy of the maximum tree height estimation using cloth simulation-based, mean-filtered and median-filtered CHMs.
The LiDAR data for the study area were collected from August to September 2012 using a Leica ALS60 system on a Yun-5 aircraft. The system wavelength was 1064 nm, and the beam divergence was 0.22 mrad. In our datasets, this system operated at a 166 kHz pulse rate at 1800 m above ground level with a field of view of approximately 30° and recorded up to four returns for each laser pulse. The scan angle statistics indicate that 97% and 99% of the pulses fall within the 0°–12° and 0°–15° zenith angles. The average point density was 8.24 points ∙ m–2 with an average spacing of 0.41 m.
Nine test plots (called L1–L9) with an area of 45 m × 45 m were established in the study area in August 2013 (Fig. 1b). The positions of the four corners of each plot were measured using a differential GPS device. The maximum tree height of each plot was measured using a hand-held laser rangefinder. The collection time between field-measured and LiDAR data was not absolutely consistent (the time interval is 1 year). Considering that the height growth rate of the tree species in the study area is relatively slow, field-measured data could be acceptable for validating the proposed algorithm (Tang 2013).
The pre-processing of LiDAR data consists of three steps, namely, noise removal, height normalisation and grid. In practice, the real and simulated data were pre-processed using the LAStools (point cloud processing software, Rapidlasso Company, Gilching, Germany) modules which include lasnoise, lasground and lasgrid (Isenburg 2017). The first step in the real data was to remove high or low noise points using the lasnoise module. Secondly, the lasground module was used to extract the ground points and then calculated the relative height above the ground for each point (i.e., z coordinate). Each point was height-normalised by replacing the height of each point with its relative height above the ground. Consequently, the height of all points classified as the ground is zero. Finally, the lasgrid module was used to grid the height-normalised points, only retaining the maximum point in each grid cell. Considering that the simulated data have no noise, and the ground height is set to zero, these data were assigned to a grid, and only the maximum point in each grid cell was retained using the lasgrid module.
Theoretical basis: cloth simulation
Comparison between different pit-free CHM construction algorithms
Mean-filtered CHM: the raw CHM was filtered using a mean filter (3 × 3 kernel).
Median-filtered CHM: the raw CHM was filtered using a median filter (3 × 3 kernel).
Evaluation of real data
Assessment of the CHM application accuracy
Maximum tree height estimation results for the three algorithms in the nine plots (unit: m)
Biases between field-measured and CHM-derived plot-level maximum tree heights for the three algorithms under different CHM resolutions (unit: m)
The CHM resolution plays an important role in estimating forest structure attributes (Chen et al. 2006; Khosravipour et al. 2014). Some optimal resolution setting principles have been proposed. For example, Nyquist sampling theory stipulates that the resolution must not be larger than half the size of the minimum object of interest (e.g., the minimum tree crown) and overly smaller than the average pulse spacing (Nyquist 1928; Chow and Hodgson 2009; Mielcarek et al. 2018). Raster CHMs with resolutions of 0.2, 0.5, 0.8, 1 and 1.5 m are constructed from the real data. This task is performed to test the robustness of the proposed algorithm for deriving a plot-level maximum tree height from CHMs under different resolutions. Table 2 shows biases between the field measurements and the maximum tree heights extracted from the CHMs constructed by different algorithms. In the mean and median filters, the resolution significantly affects the estimation accuracy of the tree height. However, our algorithm is unaffected and achieves the maximum accuracy at all resolutions.
Evaluation of simulated data
Accuracy comparison between different algorithms for filling pits (unit: m)
One of the challenges in constructing a CHM from LiDAR point clouds is the existence of data pits. In this study, we propose a novel algorithm for constructing pit-free CHMs. In this algorithm, a pit-free raster CHM is directly constructed through a cloth simulation process instead of by the combination of various steps such as pit detection, pit filling and transformation from the randomly distributed point clouds to the raster CHM. The real and simulated data are used to validate the advantages of the proposed algorithm.
The visual comparison results show that the proposed algorithm is superior to the mean and median filters. Such filters generate new values for all pixels based on their neighbourhood values, thus affecting the original structure of the canopy surface (Ben-Arie et al. 2009; Shamsoddini et al. 2013). By contrast, our algorithm only fills data pits without changing the value of any other pixels, thereby preserving the original structure of the canopy surface. The proposed algorithm provides the possibility to minimise the distortion of the canopy’s uppermost layer efficiently; this layer plays a critical role in various applications (e.g., ecological, hydrological and meteorological) sensitive to vegetation evolution at the local and regional scales (Khosravipour et al. 2014).
We also evaluate the proposed algorithm by comparing the accuracy of the tallest tree height estimation using the cloth simulation-based CHM versus the mean- and median-filtered CHMs. The positive biases for all algorithms indicate that the tallest tree height may be overestimated on the ground; this phenomenon is a possible error associated with the tree height measurement in the field (Wang et al. 2019). Our algorithm is significantly accurate at tree height estimation. The present study confirms previously published statements by Chen et al. (2017); these researchers illustrated that the mean and median filters struggle with the tallest tree height estimation because such filters smoothen the treetops based on surrounding pixels. The resolution of the cloth simulation-based CHM has no effect on the estimation accuracy of the maximum tree height. However, the estimation accuracy of the tallest tree height is sensitive to the resolutions of the mean- and median-filtered CHMs. The highest pixel value within a plot may be changed because the neighbourhood pixel values may vary with the resolution.
Furthermore, we investigate the effects of the different proportions of data pits on the accuracy of the mean and median filters and the proposed algorithm. The results show that our algorithm achieves the optimal visual performance amongst the data with different proportions of data pits. The quantitative assessment results show that the mean and median filters reduce the pit effect based on a predefined window kernel size that cannot be optimal for all different proportions of data pits. For example, the median filter can fill small pits but deteriorates large ones; this result agrees with those of previous studies (Fig. 12d) (Shamsoddini et al. 2013). The proposed algorithm accurately represents the height of the canopy surfaces, especially for point clouds with high proportions of data pits. On this basis, our algorithm can be potentially applied to process the forest LiDAR point clouds acquired under leaf-off conditions given that most points will cluster on branches or trunks rather than canopy surfaces (i.e., high proportions of data pits).
In summary, the results demonstrate that the proposed algorithm has the following advantages: (a) The proposed algorithm can effectively construct pit-free CHMs in point clouds with different proportions of data pits; (b) it can accurately fill only the pits without changing the values of all the CHM pixels, thereby preserving details of the uppermost layer of the forest canopy; and (c) it only needs to set up a CHM resolution parameter.
In this study, we presented a new pit-free CHM construction algorithm based on cloth simulation. This algorithm was applied to the LiDAR data to generate raster CHM directly. The real and simulated datasets were used to compare the performance of the proposed algorithm with those of the mean and median filters. Overall, the proposed algorithm is promising in terms of robustness and high accuracy. This characteristic is useful for practical applications. Whether our pit-free CHM can improve the accuracy of extracted tree biophysical parameters, such as biomass, is interesting to investigate in the future.
We thank Radiation Modelling and Measurement Laboratory (http://ramm.bnu.edu.cn/) for data provision. Jianbo Qi is thanked for his help with the figure preparation.
WZ, SC and XL conceived the study and contributed to the manuscript writing and editing; JS coordinated the design of the program; RH processed a part of the data; SY participated the evaluation and figure preparation. GY contributed to analyses of this study. All authors read and approved the final manuscript.
This work was supported by the National Natural Science Foundation of China (Grant Nos. 41671414, 41971380 and 41171265) and the National Key Research and Development Program of China (No. 2016YFB0501404).
Ethics approval and consent to participate
The subject has no ethic risk.
Consent for publication
All the data and relevant processing methods, including the automated generation of figures of evaluation results, belong to the host institution, namely, the Beijing Normal University.
The authors declare that they have no competing interests.
- Axelsson P (1999) Processing of laser scanner data—algorithms and applications. ISPRS J Photogramm Remote Sens 54:138–147 https://www.sciencedirect.com/science/article/abs/pii/S0924271699000088. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Ben-Arie JR, Hay GJ, Powers RP, Castilla G, St-Onge B (2009) Development of a pit filling algorithm for LiDAR canopy height models. Comput Geosci 35:1940–1949 https://www.sciencedirect.com/science/article/pii/S0098300409000624. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Chen Q, Baldocchi D, Gong P, Kelly M (2006) Isolating individual trees in a savanna woodland using small-footprint lidar data. Photogramm Eng Remote Sens 72:923–932 https://www.ingentaconnect.com/content/asprs/pers/2006/00000072/00000008/art00003. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Chow TE, Hodgson ME (2009) Effects of lidar post-spacing and DEM resolution to mean slope estimation. Int J Geogr Inf Sci 23:1277–1295 https://www.tandfonline.com/doi/abs/10.1080/13658810802344127. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Dong P (2009) Characterization of individual tree crowns using three-dimensional shape signatures derived from LiDAR data. Int J Remote Sens 30:6621–6628 https://www.tandfonline.com/doi/abs/10.1080/01431160903140761. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Gaveau DLA, Hill RA (2003) Quantifying canopy height underestimation by laser pulse penetration in small-footprint airborne laser scanning data. Can J Remote Sens 29:650–657 https://www.tandfonline.com/doi/abs/10.5589/m03-023. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Goulden T, Hopkinson C (2010) The forward propagation of integrated system component errors within airborne lidar data. Photogramm Eng Remote Sens 76:589–601 https://www.ingentaconnect.com/content/asprs/pers/2010/00000076/00000005/art00005. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Hyyppä J, Hyyppä H, Leckie D, Gougeon F, Yu X, Maltamo M (2008) Review of methods of small-footprint airborne laser scanning for extracting forest inventory data in boreal forests. Int J Remote Sens 29:1339–1366 https://www.tandfonline.com/doi/abs/10.1080/01431160701736489. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Isenburg M (2017) LAStools—efficient lidar processing software. http://rapidlasso.com/LAStools.
- Khosravipour A, Skidmore AK, Isenburg M, Wang T, Hussin YA (2014) Generating pit-free canopy height models from airborne lidar. Photogramm Eng Remote Sens 80:863–872 https://www.ingentaconnect.com/content/asprs/pers/2014/00000080/00000009/art00003. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Kraus K, Pfeifer N (1998) Determination of terrain models in wooded areas with airborne laser scanner data. ISPRS J Photogramm Remote Sens 53:193–203 https://www.sciencedirect.com/science/article/abs/pii/S0924271698000094. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Latypov D (2005) Effects of laser beam alignment tolerance on lidar accuracy. ISPRS J Photogramm Remote Sens 59:361–368 https://www.sciencedirect.com/science/article/abs/pii/S0924271605000584. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Leckie D, Gougeon F, Hill D, Quinn R, Armstrong L, Shreenan R (2003) Combined high-density lidar and multispectral imagery for individual tree crown analysis. Can J Remote Sens 29:633–649 https://www.tandfonline.com/doi/abs/10.5589/m03-024. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Lim K, Treitz P, Wulder M, St-Onge B, Flood M (2003) LiDAR remote sensing of forest. Prog Phys Geogr 27:88–106 https://journals.sagepub.com/doi/abs/10.1191/0309133303pp360ra. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Mielcarek M, Stereńczak K, Khosravipour A (2018) Testing and evaluating different LiDAR-derived canopy height model generation methods for tree height estimation. Int J Appl Earth Obs Geoinf 71:132–143 https://www.sciencedirect.com/science/article/pii/S0303243418301478. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Nyquist H (1928) Certain topics in telegraph transmission theory. Trans Am Inst Electr Eng 47:617–644 https://ieeexplore.ieee.org/abstract/document/5055024. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Popescu SC (2007) Estimating biomass of individual pine trees using airborne lidar. Biomass Bioenergy 31:646–655 https://www.sciencedirect.com/science/article/pii/S0961953407001316. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Shamsoddini A, Turner R, Trinder JC (2013) Improving lidar-based forest structure mapping with crown-level pit removal. J Spat Sci 58:29–51 https://www.tandfonline.com/doi/abs/10.1080/14498596.2012.759092. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Stereńczak K, Lisańczuk M, Erfanifard Y (2018) Delineation of homogeneous forest patches using combination of field measurements and LiDAR point clouds as a reliable reference for evaluation of low resolution global satellite data. Forest Ecosyst 5:1. https://doi.org/10.1186/s40663-017-0128-5 CrossRefGoogle Scholar
- Tang X (2013) Estimation of forest aboveground biomass by integrating ICESat/GLAS waveform and TM data. Doctoral Dissertation, University of Chinese Academy of Sciences, Beijing http://www.wanfangdata.com.cn/details/detail.do?_type=degree&id=Y2440193. Accessed 20 Apr 2019Google Scholar
- Wan P, Wang T, Zhang W, Liang X, Skidmore A, Yan G (2019) Quantification of occlusions influencing the tree stem curve retrieving from single-scan terrestrial laser scanning data. Forest Ecosyst 6:43 https://link.springer.com/content/pdf/10.1186/s40663-019-0203-1.pdf.
- Wang Y, Lehtomäki M, Liang X, Pyörälä J, Kukko A, Jaakkola A, Liu J, Feng Z, Chen R, Hyyppä J (2019) Is field-measured tree height as reliable as believed—a comparison study of tree height estimates from field measurement, airborne laser scanning and terrestrial laser scanning in a boreal forest. ISPRS J Photogramm Remote Sens 147:132–145 https://www.sciencedirect.com/science/article/pii/S0924271618303046. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Yu X, Hyyppä J, Vastaranta M, Holopainen M, Viitala R (2011) Predicting individual tree attributes from airborne laser point clouds based on the random forests technique. ISPRS J Photogramm Remote Sens 66:28–37 https://www.sciencedirect.com/science/article/abs/pii/S0924271610000651. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Zhang K, Chen SC, Whitman D, Shyu ML, Yan J, Zhang C (2003) A progressive morphological filter for removing nonground measurements from airborne lidar data. IEEE Trans Geosci Remote Sens 41:872–882 https://ieeexplore.ieee.org/abstract/document/1202973. Accessed 20 Apr 2019CrossRefGoogle Scholar
- Zhao D, Pang Y, Li Z, Sun G (2013) Filling invalid values in a lidar-derived canopy height model with morphological crown control. Int J Remote Sens 34:4636–4654 https://www.tandfonline.com/doi/abs/10.1080/01431161.2013.779398. Accessed 20 Apr 2019CrossRefGoogle Scholar
Open AccessThis article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.