Extracting Alpine Shrub Using Improved Lightweight DeepLabV3+ Network

In recent years, shrubland is an important land cover type in alpine regions, while accurate segmentation of shrubs using remote sensing data remain challenging. To address these issues, this study proposes an alpine shrub segmentation method based on an improved lightweight DeepLabV3+ network, in which MobileNetV2 is used to replace the original backbone to reduce model complexity while maintaining feature representation capability, a channel squeeze-and-excitation (cSE) attention module is introduced to enhance the response to key shrub features and boundary details, and Ghost convolution is incorporated to reduce computational redundancy while preserving segmentation accuracy. Experimental results from both ablation and comparative studies demonstrate that the proposed model achieves a mean intersection over union (MIoU) of 88.47%, mean pixel accuracy (mPA) of 92.93%, F1-score of 91.80%, and overall accuracy of 94.52%, representing improvements of 3.53%, 2.64%, 2.96%, and 1.69%, respectively, over the original DeepLabV3+ model, while also significantly reducing the number of parameters and model size. In addition, independent cross-year validation using unmanned aerial vehicle (UAV) imagery acquired in 2025 suggests that the proposed model has good applicability under similar UAV sensor and acquisition conditions. Overall, this study provides an effective lightweight semantic segmentation approach for alpine shrub segmentation from high-resolution UAV imagery and offers useful technical support for vegetation monitoring in alpine regions such as the Qinghai–Tibet Plateau.