Abstract: This study presents a novel deep learning (DL) model for tropical forest canopy mapping using high-resolution multisource remote sensing (RS) data. The model integrates soft attention mechanisms and recurrent residual (R2) networks with a U-Net backbone to enhance feature focus and capture contextual information. Evaluated using the mean intersection over union (MIo U) metric, the proposed model achieved 0.9521 accuracy, surpassing the traditional U-Net (0.9266), U-Net with attention (0.9383), and Deeplabv3+ (0.9419) models. It also generated superior forest canopy change maps. These improvements enable better monitoring of tropical forests, which face increasing anthropogenic threats from deforestation, urban sprawl, agriculture, and climate change.