CLASSIFICATION OF TREES IN HYPERSPECTRAL CANOPY DATA USING MACHINE LEARNING: COMPARATIVE ANALYSIS OF FOREST STRUCTURE COMPLEXITY

Abstract

The classification of tree species by remote sensing is an important task with a broad range of applications, including forest management, environmental monitoring, and climate change studies. Hyperspectral imaging has proven to be a valuable tool for this classification. Additionally, deep learning techniques have obtained outstanding results in hyperspectral classification. In this study, we apply a neural network to the classification of aerial hyperspectral images of trees. The study was conducted at a research station in southern Chile with 32 tree species. Our database has 3080 tree canopies that have been manually segmented and classified. The goal of the work was to study the correlation between forest structure complexity and classification performance across three different forest conditions: native forest, plantation of native species, and plantation of exotic species. The results show that classification performance is higher when forest structure and composition are simpler. We used a ResNet neural network as classifier and compared its performance with support vector machine and random forest. The best performance was obtained using ResNet in exotic plantations, the forest condition with the simplest structure, achieving an F1-score of 85.23%.