Abstract: Arid valleys in the upper reaches of the Minjiang River represent a typical ecologically fragile region across China. Accurately delineating their boundaries holds great significance for regional ecological restoration and hazard risk assessment. Based on multi-source remote sensing data, this study conducted boundary delineation of the arid valleys using three semantic segmentation models: U-Net, DeepLabV3+, and PSPNet. To mitigate the class imbalance caused by an extremely low proportion of valley pixels, a hybrid loss function combining Dice loss with Focal loss was incorporated to enhance the models’ capability to identify boundary details. The results indicate that the DeepLabV3+ model outperformed its counterparts, yielding a mean intersection over union (mIoU), precision, recall, and Dice coefficient of 83.99%, 91.20%, 90.01%, and 90.59%, respectively. The atrous spatial pyramid pooling (ASPP) module of the DeepLabV3+ model can effectively capture the multi-scale characteristics of the arid valleys, rendering this model highly suitable for automated boundary delineation. Time-series analysis based on this model indicated that during 2013—2025, the total area of the arid valleys reduced by 27.18% and exhibited a three-stage evolutionary pattern consisting of slow expansion, rapid shrinkage, and gradual stabilization sequentially.