Rapid and intelligent sugarcane identification: a comparative study of multiple deep learning models

The complex terrain and persistent cloud cover in karst agricultural regions pose significant challenges for rapid and accurate sugarcane identification. To address these challenges, this study integrates optical and microwave remote sensing data and employs four representative deep learning models—MVT-UNet, U-Net, DeepLabV3+, and SegFormer—to perform intelligent sugarcane identification experiments in Leiping Town. The performance of these models was systematically evaluated in terms of segmentation accuracy and computational efficiency. Experimental results show that all four models can effectively accomplish sugarcane segmentation tasks. Among them, U-Net achieved the highest segmentation accuracy, while MVT-UNet and DeepLabV3+ also demonstrated strong and stable performance. Although SegFormer exhibited a clear advantage in training speed, its segmentation accuracy was comparatively lower. This study provides a comprehensive comparative evaluation of multiple deep learning architectures for crop identification in karst agricultural environments, offering valuable insights for the protection of the sugarcane industry and the sustainable development of regional agriculture.