The Yuanjiang dry-hot valley is a typical severe gully erosion region in southwest China. However, due to the steep terrain of canyon areas, implementing traditional field contact-based monitoring methods is challenging, which highlights the high potential of deep learning-based automatic gully monitoring. The complex morphology and small area proportion are the key limiting factors in automatic gully mapping using deep learning methods. This study established a gully classification strategy that distinguishes between gully with clear boundary lines (GC) and unclear boundary lines (GU) across 3 representative locations characterized by high heterogeneity in soil properties and gully size/morphology. Five deep learning models were trained on distinct gully-type datasets to evaluate this framework. The results revealed that SegFormer achieved the highest Precision (69.67%) when mapping undifferentiated gullies. For GC, Twins exhibited the optimal performance (72.16% Precision), whereas DeepLabV3+ achieved the best Precision (59.47%) for GU. These findings confirm that exclusive training on GC gullies yields higher mapping accuracy, whereas detecting GU gullies requires incorporating GC samples to optimize performance. Furthermore, this study demonstrates that low gully area proportion adversely affects the mapping accuracy of deep learning-based models and proposes a pixel-balancing mechanism - specifically, inverse frequency weighting based on pixel quantity - to mitigate this systematic bias, resulting in an average Precision improvement of 7.12%. This integrated approach provides a methodological foundation for large-scale gully erosion mapping in similar extreme environments.