To address the challenges of crop semantic segmentation in complex intercropping agricultural scenarios, this study proposes a Transformer network based on multi-scale feature enhancement. The research aims to overcome the insufficient segmentation accuracy of conventional methods in intercropping environments characterized by large crop scale variations and strong weed interference. The proposed approach incorporates two innovative modules. First, a multi-scale pooled self-attention module employs a parallel multi-scale pooling strategy, combined with channel reduction and feature fusion, to extract rich multi-scale contextual information. Second, a cross-spatial feed-forward network module introduces local and channel interaction mechanisms, enhancing the modeling of spatial dependencies through grouped attention. Experiments conducted on intercropping datasets from Alaer and Bachu in Xinjiang demonstrate that the proposed method achieves 80.5% mIoU and 87.7% mAcc on the Alaer dataset, and 78.8% mIoU and 86.9% mAcc on the Bachu dataset, significantly outperforming mainstream comparative models. Ablation studies verify the effectiveness of the proposed modules, with the complete model achieving a 4.7% mIoU improvement over the baseline Transformer while maintaining a relatively low computational cost of 15.9M parameters and 31.1G FLOPs. Visualization results further confirm the model’s adaptability to multi-scale crops, maintaining an inference speed of 38–40 f/s at different flight altitudes. This study provides an effective solution for accurate segmentation in complex intercropping scenarios and offers substantial theoretical significance and practical potential.
Wang et al. (Wed,) studied this question.