CAM-HRNet: A Multi-Scale Parallel Structure Change Detection Framework for Forest Land

As a vital component of Earth’s ecosystem, forest land plays an irreplaceable role in maintaining ecological balance. To address the limitations of traditional deep learning methods in remote sensing data generalization, this paper proposes CAM-HRNet, a change detection framework that integrates the High-Resolution Net (HRNet) backbone with a convolutional attention mechanism. The method leverages HRNet’s multi-scale parallel structure and incorporates attention mechanisms at different scales, while a step-by-step up-sampling mechanism enables smooth multi-scale feature fusion with reduced information loss. By combining CAM-HRNet’s feature extraction capabilities with a twin network structure, we develop both binary and semantic change detection methods. The binary method identifies changed areas through difference feature calculation, while the semantic method adds segmentation branches to analyze specific change types. Focus loss is introduced to address sample imbalance. Experimental results demonstrate that our binary change detection achieved 75.24% IoU and 85.88% F1, outperforming comparable methods by 2.62% and 1.74%, respectively, while effectively filtering irrelevant changes. The semantic method maintained strong performance with 63.93% IoU and 77.53% F1, providing accurate monitoring of forest ecosystem dynamics to support resource protection and management.