Deforestation Segmentation Approach Based on Time of Event Occurrence Using Multitemporal Satellite Data

Abstract

Deforestation in the Amazon rainforest is approaching historically worst levels, and accurate estimation of deforested areas is crucial to protect against further deforestation. In this letter, we propose a forest deforestation segmentation approach based on Mask2Former, which utilizes multisatellite information and an optimized backbone network. Particularly, we enable the determination of whether to use the corresponding optical imagery during the training phase based on the amount of cloud cover and applied a model to compensate for the resulting data sparsity. The model reflects deforestation occurrence information at specific points in time in the deforestation predictions at points in the same time series, thereby detecting deforestation with high accuracy while compensating for data shortages. In experiments, the proposed method achieved excellent performance, with a pixel accuracy of 91.1% and an F1 score of 88.8%. This method was validated by achieving the best segmentation performance for deforested areas in the official CVPR MultiEarth Workshop 2023 challenge.