Abstract: Deep learning offers an efficient approach to forest land change detection by automatically extracting change features from dual-time-phase high-resolution images through large-scale data training, reducing reliance on operator experience and enhancing detection efficiency. Focusing on Zhejiang Province, this study utilized high-resolution remote sensing images from World Imagery Wayback and applied deep learning models, including UNet series, DeepLabV3 series, FCN, and SegNet, for binary classification of forest land changes. All models demonstrated strong performance on the test set, with mIoU above 84.70%, Accuracy above 92%, F1-score above 91.00%, and Recall above 91%. Among them, the UNet model achieved the best results, with mIoU of 89.54%, Accuracy of 95.15%, F1-score of 94.44%, and Recall of 94.13%, though all models showed varying degrees of edge contour blurring and incomplete intra-class connectivity. To address these limitations, the UNet model was improved using attention mechanisms, including SE (Squeeze-and-Excitation), SGAM (Spatial Gated Attention Mechanism), and CBAM (Convolutional Block Attention Module), and similar enhancements were applied to DeepLabV3 and DeepLabV3P models for control experiments. Results showed that the UNet+SE model, leveraging channel attention, achieved the greatest improvement in mIoU (0.18%), while models enhanced with spatial attention (UNet+SGAM, DeepLabV3+SGAM, DeepLabV3P+SGAM) experienced decreases in mIoU by 1.01%, 0.77%, and 0.67%, respectively, indicating that channel features are more critical than spatial features for forest land change detection. These findings confirm the UNet model’s superior performance and highlight the importance of prioritizing channel features, providing valuable insights for forest land change detection using deep learning and high-resolution remote sensing imagery.