Improved lightweight DeepLabV3+ for bare rock extraction from high-resolution UAV imagery

Bare rock information extraction in karst regions is crucial for geological hazard monitoring and ecological assessment. However, in sparsely vegetated areas, bare rock exhibits similar spectral characteristics to surrounding land cover, and the boundaries are often indistinct, making it challenging for traditional classification methods to distinguish these transitional zones accurately. To address these challenges, this study proposes a bare rock extraction method based on an improved lightweight DeepLabV3+ model. MobileNetV2 is used as the backbone network, and the Channel Attention Module (CAM) and Spatial Attention Module (SAM) are introduced to enhance feature extraction capability. Results show the following: (1) When MobileNetV2 is used as the backbone of DeepLabV3+, the Accuracy, F1 score, and MIoU reach 97.39 %, 78.91 %, and 82.11 %, respectively, outperforming VGG16, Xception, SqueezeNet, and traditional segmentation models. (2) Applying the lightweight DeepLabV3+ model to bare rock identification in orthophoto imagery of the study area results in a bare rock rate error of approximately 5 %, demonstrating the practical applicability of the model. (3) After the introduction of the attention mechanism, the model's Recall, F1 score, and MIoU increased by 14.00 %, 8.37 %, and 5.62 %, respectively, remarkably enhancing identification completeness and boundary accuracy. Meanwhile, the improved model had a parameter count of 6.98 M and a computational complexity of 7.24G, achieving enhanced accuracy while maintaining computational efficiency. The research results can provide accurate bare rock information to support geological hazard monitoring and early warning, and offer new technical solutions for ecological restoration and risk assessment. (Data sets and code links: https://figshare.com/articles/dataset/Bare_rock_dataset/28143443?file=53186633 ). • A lightweight DeepLabV3+ model is improved to mitigate confusion in bare rock identification. • CAM and SAM modules were introduced to enhance the model's ability to identify bare rock edges. • While maintaining a lightweight structure, the improved model achieved an MIoU of 87.73 %, outperforming traditional models and the TransUNet. • A new dataset for bare rock identification from high-resolution UAV imagery is introduced.