Abstract: In response to the problem of abnormal segmentation in farmland, a feature fusion skip connection module and a global-local attention module were used to improve the UNet network model. A real-time segmentation network for abnormal farmland areas was proposed, which achieved fine segmentation of various abnormal farmland areas. The results showed that the Mean Intersection Union ratio （MIoU） of the real-time segmentation model for abnormal growth areas in farmland was significantly better than that of other models, with a MIoU of 41.24%;compared to the model using UNet as the baseline, although the number of parameters in this study model had slightly increased, the farmland segmentation effect had significantly improved, with an increase of 4.16 percentage points in MIoU;compared with the SegFormer model based on Transformer encoder, the parameter count of this study model was basically the same, with an increase of 2.50 percentage points in MIoU. This research model ensured excellent segmentation performance in each category by using adaptive sampling training methods. Using multispectral aerial images to train a real-time segmentation model for abnormal growth areas in farmland could help achieve real-time monitoring and early warning of farmland growth by drones, promote the development of smart agriculture, and provide new methods and ideas for automatic monitoring of farmland growth.

Key words: multispectral, abnormal growth areas in farmland, aerial images, UNet, DeepLabV3+, SegFormer