DTFD : A Transformer Approach for High‐Resolution Satellite Image Forest Change Detection

ABSTRACT Deforestation is a significant contributor to global greenhouse gas emissions, underscoring the need for effective forest conservation and management strategies. Developing such strategies requires a thorough understanding of the primary drivers of forest loss. However, the complexity of these factors, combined with the requisite skill set for accurate identification, poses considerable challenges for data collection. This study introduces a novel deep learning‐based approach, termed Deep Transformation Forest Detection (DTFD), which utilizes vision transformers equipped with a self‐attention mechanism. This innovative method enhances the modeling of contextual and spatial relationships in satellite imagery while facilitating efficient processing without relying on convolutions. This capability is particularly beneficial for heterogeneous and binary classification tasks. The self‐attention mechanism allows for the assignment of varying weights to input data, thereby improving the identification of areas at risk of deforestation adjacent to forested regions. The results achieved by DTFD demonstrate exceptional performance compared to state‐of‐the‐art methods across multiple datasets. Notably, the findings reveal significant changes in forest cover and environmental dynamics, with DTFD attaining superior metrics, including accuracy (95.64%), precision (95.55%), F 1 score (93.74%), recall (94.83%), and Intersection over Union (IoU) (94.31%). This research contributes to the monitoring of climate change impacts, rapid urbanization, and natural disasters, with a specific emphasis on urban forests and their interactions with urban environmental changes.