A Full-Spectrum Time-Reversal Deep Network for Label Imbalance Learning: A Case Study With Novel Ground-Truth Labeling Strategy for Mangrove Forest Change Detection

Mangrove forests act as powerful carbon sinks as they store much larger amounts of carbon than terrestrial forests, and hence help mitigate climate change, echoing the sustainable development goals. Mangrove forest change detection (CD) technique, monitoring the mangrove gains and losses over time, has drawn considerable attention in the optical satellite remote sensing area in recent years. CD is theoretically more challenging than the mangrove mapping at a single time instance, because it requires accommodation of different illumination and atmospheric conditions across time. Another challenge lies in the ground-truth labeling, as the spatial resolution of satellite images often necessitates manual inspection to establish accurate labels. To address these obstacles, we develop a time-reversal feature learning (TRFL) to leverage the existing training samples effectively, and to force the network to output explainable CD results (i.e., when the bitemporal network inputs are interchanged, forward gains should become backward losses, and vice versa), thereby increasing the reliability/stability with more logical inference. Remarkably, mangrove losses are much more frequently seen than mangrove gains, making the training labels quite imbalanced, while the TRFL virtually creates more data corresponding to the mangrove gains, thereby mitigating the labeling imbalance. As a side contribution, we propose a novel yet effective strategy for labeling the mangrove areas and changes. Comprehensive experiments demonstrate that TRFL achieves state-of-the-art performance in dynamic CD tasks for mangrove forest analysis.