DeepForMaP

Over the last decade, the Earth has witnessed significant climate change and global warming, largely as consequences of rapid industrial development. To combat this, the management and preservation of forests have become increasingly essential, given their homeostatic role in regulating the global climate. Point clouds (PCs) obtained from LiDAR mapping offer vast amounts of data, which have proven invaluable for forest mapping. In our research, DeepForMaP, we leverage PC data and utilise advanced deep learning techniques to monitor forest patterns across both space and time. One of the primary challenges in processing PC data arises from its unstructured nature. To address this, we employ recent feature extractors such as Mamba and Kolmogorov Arnold Networks (KANs) to process the data, extracting meaningful and discriminative representations that result in more accurate classification and detection. Furthermore, given the high point density of PCs (often in billions), data annotation is limited. To overcome this, we employ self-supervised learning and zero-shot learning techniques. The former capitalises on the inherent structure of PCs to design specific tasks that train feature extractors, while the latter leverages semantic information from the corresponding PC descriptions to manage data scarcity. Our trained models are applied to classification and change detection tasks, and we also evaluate their performance in cross-domain settings, where training and testing occur in geographically disjointed areas. This project is expected to result in efficient deep learning models, deployable in practical forest mapping scenarios. DeepForMaP will be hosted at IRISA laboratory at Université Bretagne Sud Vannes under the esteemed supervision of Prof. Sébastien Lefèvre and Dr. Minh-Tan Pham. This will offer the candidate invaluable exposure to cutting edge scientific development in earth observation and and equip him to become an independent researcher in the future.