Forest-Type Change Detection by Synergizing Multi-Source Remote Sensing Data in a Deep Learning Model FCCDNet

This study introduces an end-to-end deep learning model (FCCDNet) for forest-type change detection using bi-temporal Sentinel-1 (S1) or Sentinel-2 (S2) images. FCCDNet consists of a parallel Swim Transformer backbone network, a feature aggregation module, and a multitask learning decoder. The trained model can produce maps highlighting the areas with a change on forest types, using a pre-event and post-event image pair, either S1-S1, S2-S2, S1-S2, or S2-S1. The performance was validated and demonstrated in southern Sweden between 2018 and 2023. FCCDNet achieved land cover classification accuracy of 93.26% and change detection accuracy of 90.56% when using S2-S2 pair, significantly outperforming four other algorithms tested. When using S1-S2, S2-S1, or S1-S1 pairs, changes can also be detected but with lower accuracy (65.94% - 76.68%). The results show that FCCDNet can achieve forest change detection with higher accuracy than commonly used methods, and can tolerate certain levels of replacement from multispectral images to SAR images in cloudy conditions. The FCCDNet model also showed robustness against salt-and-paper effects of mapping and exhibited sensitivity to changes with smaller amplitude. The FCCDNet model and data processing framework could support the dynamic monitoring of forest resources with high automation and fast response.