Deep learning methods for modelling forest biomass and structures from hyperspectral imagery

Forests affect the environment and ecosystems in multiple ways. Hence, understanding the forest processes and vegetation characteristics help us protect the environment better, reserve the biodiversity, and mitigate the hazardous impacts of climate change. There are studies in hyperspectral remote sensing that employ both empirical and artificial intelligence (AI) methods to analyze and predict the vegetation parameters. However, these methods have weaknesses. First, the empirical methods are inefficient because they cannot fully utilize a large amount of hyperspectral data. Secondly, even though the existing AI-based methods can achieve remarkable results, they are only validated on small-scale datasets that have simple forest structures. Thus, a robust technique that can effectively model complex forest structures on large-scale datasets is an open challenge. This thesis directly addresses the challenge by proposing a novel deep learning architecture that can jointly learn and model four discrete and twelve continuous forest parameters. The final model is comprised of three 3D convolution layers, a 3D multi-scale convolution block, a shared fully-connected layer, and two fully-connected layers for each learning task. The model uses a loss, namely focal loss, to address class imbalance problem and the gradient normalization for multi-task learning. Then, we record and compare the results of our comprehensive experiments. Overall, the proposed model reaches 78.32% class-balanced accuracy for the four classification tasks. For the regression tasks, the model achieves a notably low average mean absolute error (0.052) and high Pearson correlation coefficient (0.9) between predicted and target labels. In the end, the shortcomings of the thesis work are discussed and potential research areas for future work are suggested