ForResANeXt: Forest/non-forest segmentation with aggregated residual attention network in satellite imagery

Accurate mapping of forest (F) and non-forest (NF) areas is essential for ecological assessment, resource management, and deforestation monitoring. However, complex backgrounds, severe class imbalance and redundant features continue to limit the accuracy and efficiency of network segmentation. To overcome these issues, we present ForResANeXt, a novel semantic segmentation network that uses Sentinel-2 multispectral imagery for forest/non-forest mapping. The model incorporates an AResCAB to enrich contextual feature representations while reducing redundancy and a lightweight embedded attention module to improve positional awareness. Furthermore, attention-gated skip connections suppress background noise and emphasize key spatial information, and a Focal Dice Loss function mitigates the impact of severe class imbalance. Experimental results demonstrate that ForResANeXt achieves a mIoU of 95.31%, surpassing U-Net and mainstream CNN variants in recall and F1 score for the minority non-forest class. It also outperforms several representative advanced CNN architectures and Transformer-based models in terms of Boundary IoU and Small Object Recall. Qualitative comparisons further confirm its superior capability in preserving structural details and delineating complex boundaries with reduced misclassification. Cross-regional transfer experiments validate the model’s robustness and generalization capability across diverse geographical and temporal conditions, and ablation studies confirm the effectiveness of each proposed component. Overall, ForResANeXt shows great promise for efficient and accurate forest cover mapping using multispectral satellite data.