Abstract: Tanking Qianzhong region as study area and leveraging prior knowledge of cultivated land from Third National Land Survey database, effectiveness of DeepLabV3, MMSegmentation, and U-Net in extracting non-grain conversion information from GF-2 satellite imagery was comparatively analyzed, to provide insights for identifying non-grain conversion in mountainous cultivated areas．Results indicated that U-Net model achieved the highest overall accuracy during training phase, with recall, precision, and F1-score all reaching 95%．MMSegmentation followed, while DeepLabV3 performed comparatively low．However, DeepLabV3 model demonstrated a 75% improvement in computational efficiency than U-Net, with significantly faster training and classification extraction speeds．All three methods attained non-grain conversion monitoring accuracies exceeding 85%．Among them, DeepLabV3 showed the highest precision, with an intersection over union of 86.22%, and an overall accuracy of 94.27%．A total of 2038 non-grain conversion patches were extracted in study area, covering 235.15 hm²．These patches exhibited a spatial distribution characterized by "scattered large patches and concentrated small patches"．Non-grain conversion incidence in the dam area reached as high as 2.96%, while that in mountainous cultivated areas was only 0.44%, indicating greater non-grain conversion intensity in flatlands than in mountainous areas．