Abstract: Background Slope farmlands are widely distributed in the alpine and canyon regions of Southwest China, characterized by fragmented complex landforms and fragile ecological environments. Efficient and accurate extraction of the spatial distribution information of slope farmlands is critical for soil erosion control and agricultural resource monitoring of slope farmlands. This study proposes a deep learning-based intelligent interpretation method integrating multi-source features to realize efficient and precise extraction of slope farmlands in alpine and canyon areas of Southwest China. Methods High-resolution satellite remote sensing images and Digital Elevation Model (DEM) data were fused to construct a multi-band fused feature dataset, which strengthens the discriminative support of topographic features for slope farmland identification. A systematic comparative evaluation of comprehensive performance was carried out on 14 semantic segmentation models. The DeepLabV3Plus model with optimal overall performance was selected via the relative sum of gaps method, and systematic tuning was conducted on its key hyperparameters including optimizer and batch size. A morphological post-processing and vectorization module for slope farmland patches was developed based on C++17, forming a fully automated workflow for intelligent interpretation of slope farmlands in Southwest China’s alpine and canyon regions. Results The DeepLabV3Plus model achieved the optimal comprehensive performance with the minimum relative sum of gaps of 0.3399. The optimal hyperparameter combination adopted the Adam optimizer, a batch size of 16, and a cosine annealing learning rate scheduler, balancing interpretation accuracy and computational efficiency. The average overall accuracy of model validation reached 74.85%, with average precision and recall of 78.4% and 70.4%, respectively, and an average F1-score of 73.1%. The processing time for a single image tile (256×256 pixels) was approximately 10 ms, demonstrating great potential for large-scale engineering applications. The introduction of DEM topographic features significantly improved the model’s discrimination capability in complex mountainous areas and effectively alleviated the problems of "same object with different spectra" and "same spectrum with different objects", making the method particularly suitable for intelligent interpretation and extraction of slope farmlands in topographically variable alpine and canyon regions of Southwest China. Conclusions This study develops a novel workflow balancing accuracy, efficiency and operability for intelligent interpretation and extraction of slope farmlands in high-relief, fragmented alpine canyon terrain. The research outcomes can provide reliable technical support for special surveys of slope farmlands, soil erosion governance, and dynamic monitoring of agricultural resources.