Geodaysit 2023

Autores: Abderrahim Nemmaoui, Fernando J. Aguilar, Manuel A. Aguilar
Forests act as important carbon sinks, therefore being key components of the global carbon cycle. The carbon dioxide emissions account is essential for climate regulation policies and the evaluation of the effects of these policies, as well as for understanding the services they provide to societies.
Traditionally, forest inventories are completed by ground-based expert crews. These field surveys are uneconomical, time consuming and not adequate for studies dealing with periodic data collection. Consequently, one of the key topic in forest applications is to find an effective method to produce effective and accurate inventories.
In recent years, Remote Sensing (RS) has proven to be capable of providing independent, timely and reliable forest information. RS data are used to estimate several forest variables of silvicultural interest such as crown diameter (CD), tree height (H), diameter at breast height (DBH) and aboveground biomass (AGB). In this sense, and due to its ability to estimate attributes at tree level, LiDAR derive point cloud data has become a valuable data source in the field of efficient and accurate detection and segmentation of individual trees (IT).
State-of-the-art approaches use different algorithms for individual tree segmentation (ITS). For each algorithm, a specific methodology to create the input Canopy Height Model (CHM) and/or many parameters should be tuned to somehow adapt the segmentation algorithm to each particular forest stand. This approach makes the results highly dependent on the applied local fitting parameters, which implies difficulties when applied for large-scale mapping. In addition, the parameter setting process is quite time consuming and requires learning and understanding the meaning and role of each parameter.
The main goal of this work aims at developing a pipeline that requires minimal user interaction when working on large areas of Mediterranean forests. The expected results should facilitate the production of broad-extend IT maps and extract the corresponding dendrometric parameters from low-density airborne laser scanning (ALS) data without spending time tuning algorithm parameters.
The study area was located in Sierra de María-Los Vélez Natural Park (Almeria, Spain). Up to 38 reference square plots of 25 m side containing reforested stands of Aleppo pine (Pinus halepensis Mill.) with variable density, tree height and presence of shrubs and low vegetation mainly represented by little holm oak trees (Quercus ilex L.). This forest composition and structure make up a forest typology that is very representative of the Mediterranean forests.
Three open source raster-based (i.e., CHM-based) were tested to extract tree location and some dendrometric parameters such as tree H and CD. The first algorithm is the method proposed by Dalponte & Coomes(2016) adapted and introduced in the package lidR (Roussel et al.2020). The second one is the algorithm developed by Silva et al.(2016), which is focused on the way to better approximating the intersecting canopy of multiple trees after locating treetops by local maxima. The last algorithm tested is included in the library Digital Forestry Toolbox (DFT). In addition, the point cloud-based algorithm proposed by Li et al.(2012) was also tested.
For every algorithm tested, we tried different parameters to find the best pipeline, finally obtaining up to 4024 combinations of all tested algorithms for each experimental plot. For each setting, tree detection accuracy was assessed by computing the detection rate, and the commission and omission errors. Some statistics, such as median, RMSE and relative RMSE, were also used to quantitatively assess the accuracy of tree H and CD estimates over each reference plot.
The IT detection accuracy rates, in terms of precision, recall, and F1-score, showed the successful performance of the pipeline proposed in this study. The algorithm proposed by Li et al.(2012) showed detection F1-score average values of 82.65% (using the same parameter combination for the 38 experimental plots). However, it failed in delimiting the crown diameter (relative RMSE 57.06% and Pearson r of 0.55). The method developed by Silva et al.(2016), when applied on a CHM generated with the point-to-raster algorithm and using a LM based on a variable Tree Window Size (TWS), presented a similar F1-score for ITS (i.e., 82.53%), but being most successful delimiting the crown (relative RMSE 22.21% and Pearson r of 0.68). Finally, Dalponte & Coomes(2016) and DFT methods showed slightly worse results, with average F1-scores of 80.41% and 75.66%, respectively.
The results obtained confirms the usefulness of low-density ALS data to both detect IT and estimate H and CD, also underlining some key aspects regarding the choice of the correct method and parameters to perform single tree detection for Aleppo pine in large areas of Mediterranean forests.