Towards forest sustainability : detecting and characterizing forest tree decay levels from airborne LIDAR and CIR aerial imagery using deep learning and LiDAR metrics

Abstract

Forests serve as a key component for life on Earth. This is why forests are the central focus of the UN Sustainable Development Goal (SDG15). Nonetheless, forests are subject to disturbances due to natural and anthropogenic activities that lead to forest tree mortality. This results in less productivity and ecosystem services. Thus, sustainable forest management becomes a necessity, in order not to render the forest extinct. Forest tree decay is a forest mortality phenomenon, thus, detecting, classifying, and characterizing forest tree decay levels will be significant for understanding forest growth dynamics and health, for effective management. This will equally be useful for biomass, biodiversity, and carbon stock assessments. However, for sustainable forestry, an automated approach is required for cost-effectiveness. Aiming toward forest sustainability, this thesis focused on detecting and characterizing forest tree decay levels from airborne LiDAR and CIR aerial imagery using deep learning (DL) and LiDAR metrics. Consequently, a consolidation of coherent investigations was conducted toward the aim. First, this thesis presents an insightful review analysis and evaluation of existing works on DL applications in remote sensing-based forest health studies and establishes the nonexistence of studies related to forest tree decay levels. This investigation revealed that most of the studies modified existing state-of-the-art DL algorithms to reach improved results, while LiDAR data application was very less. Adding to knowledge, this research gives insight into the scientific advances of DL in remote sensing-based forest health studies and perspectives for future development, for stakeholders. Secondly, this thesis presents an approach to an improved Mask-RCNN model for instance segmentation of standing dead trees from multispectral aerial imagery. Despite a limited training sample, the model reached an Average Recall, F1-Score, and Mean Average Precision (mAP) of 88%, 87%, and 85% respectively. While this investigation demonstrates the significance of aerial imagery and DL for enhanced forest management, it also indicates the limitation of imagery for forest tree decay level classification. Next, the aerial imagery was augmented by combining ALS and CIR imagery for five forest tree decay levels classification based on two ML/DL approaches; 2D-based using CNN, and Random Forest, and 3D-based using KPConv. This model achieved OA reaching 88.8%, 88.4%, and 85.9% for KPConv, CNN, and RF respectively. This study shows the significance of LiDAR and sensor combination in ML/DL-based forest tree decay level classification. Finally, for the first time, LiDAR metrics through estimations of LAI/LAD and L-moment ratios were used to characterize and model forest plot decay levels from ALS data. Two bivariate models showed that 96% (R2 = 0.96) and 92% (R2 = 0.92) of the variation in Gap Fraction (GF) and third L-moment ratio (T3) respectively, were explained by the Vegetation Area Index (VAI) with a significant negative association, at a 95% confidence interval. Furthermore, upon testing the rule-based method, the forest plot decay levels were classified into two categories corroborating the concepts of Euphotic and Oligophotic forest areas.