Identification of Significative LiDAR Metrics and Comparison of Machine Learning Approaches for Estimating Stand and Diversity Variables in Heterogeneous Brazilian Atlantic Forest

Abstract

Data collection and estimation of variables that describe the structure of tropical forests, diversity, and richness of tree species are challenging tasks. Light detection and ranging (LiDAR) is a powerful technique due to its ability to penetrate small openings and cracks in the forest canopy, enabling the collection of structural information in complex forests. Our objective was to identify the most significant LiDAR metrics and machine learning techniques to estimate the stand and diversity variables in a disturbed heterogeneous tropical forest. Data were collected in a remnant of the Brazilian Atlantic Forest with different successional stages. LiDAR metrics were used in three types of transformation: (i) raw data (untransformed), (ii) correlation analysis, and (iii) principal component analysis (PCA). These transformations were tested with four machine learning techniques: (i) artificial neural network (ANN), ordinary least squares (OLS), random forests (RF), and support vector machine (SVM) with different configurations resulting in 27 combinations. The best technique was determined based on the lowest RMSE (%) and corrected Akaike information criterion (AICc), and bias (%) values close to zero. The output forest variables were mean diameter at breast height (MDBH), quadratic mean diameter (QMD), basal area (BA), density (DEN), number of tree species (NTS), as well as Shannon–Waver (H’) and Simpson’s diversity indices (D). The best input data were the new variables obtained from the PCA, and the best modeling method was ANN with two hidden layers for the variables MDBH, QMD, BA, and DEN while for NTS, H’and D, the ANN with three hidden layers were the best methods. For MDBH, QMD, H’and D, the RMSE was 5.2–10% with a bias between −1.7% and 3.6%. The BA, DEN, and NTS were the most difficult variables to estimate, due to their complexity in tropical forests; the RMSE was 16.2–27.6% and the bias between −12.4% and −0.24%. The results showed that it is possible to estimate the stand and diversity variables in heterogeneous forests with LiDAR data.

Enhancing forest attribute prediction by considering terrain and scan angles from Lidar point clouds: a neural network approach

“…A few studies have explored different deep-learning methods to predict forest attributes from lidar data. [40] used an MLP architecture with the principal components of a set of metrics similar to our study. They observed an rRMSE of 22.5% for the predictions of BA in heterogeneous tropical forests, but with gentle or no relief.…”

“…In our study, despite the mountainous relief, which is known to add issues in ABA modelling, the bestperforming model was the expterrain+scan dataset with an rRMSE of 19.9%. However, [40] did not consider metrics such as the gap fraction with proven explanatory power for forest structure characterisation. Additionally, the gap fraction and the rumple index (used in this study) are metrics sensitive to lidar scan angle [33], [41].…”

“…The choice of lidar metrics is also a part of the modelling strategy. We used the intensity information provided by the data provider without implementing a normalisation step to enhance it, as demonstrated in different studies [40], [42], [43]. [44] showed that intensity information is more effective than heightbased metrics in discriminating tree species.…”

Enhancing Forest Attribute Prediction by Considering Terrain and Scan Angles From Lidar Point Clouds: A Neural Network Approach

“…A few studies have explored different deep-learning methods to predict forest attributes from lidar data. Martins-Neto et al [40] used an MLP architecture with the principal components of a set of metrics similar to our study. They observed an rRMSE of 22.5% for the predictions of BA in heterogeneous tropical forests, but with gentle or no relief.…”

“…In our study, despite the mountainous relief, which is known to add issues in ABA modeling, the best-performing model was the exp terrain+scan dataset with an rRMSE of 19.9%. However, Martins-Neto et al [40] did not consider metrics such as the gap fraction with proven explanatory power for forest structure characterization. Additionally, the gap fraction and the rumple index (used in this study) are metrics sensitive to lidar scan angle [33], [41].…”