Segmentation of wheat farmland with improved U-Net on drone images

Accurate farmland segmentation is essential for modern agriculture and automated navigation. We propose an improved U-Net for farmland area segmentation. The wheat farmland data images were collected at the winter wheat experimental base of the Institute of Agricultural Economics and Information, Henan Academy of Agricultural Sciences. U-Net adopts the encoder–decoder structure and skips connection to achieve segmentation. The downsampling operation in the encoder stage weakens the detailed features. The semantic gap between the decoder and the encoder will cause the sparse wheat seedlings in the farmland cannot be captured. Based on the above problems, the improved U-Net uses a multiscale global attention module (MGA) in the bottleneck layer. MGA forms enhanced features by aggregating multiscale global context information and using an improved attention mechanism. An interaction mechanism (IM) is added between the decoder and the encoder. The encoder–decoder IM concatenates multiple attention units and fuses them with the original features on the encoder side to update the input features to the encoder. To lighten the model, we also define two multiplexed convolution kernel sequences in the code, which are shared by all encoders or decoders. The method proposed in this paper is evaluated on the farmland segmentation dataset. Significantly better segmentation results are achieved compared to classical models (U-Net, U-Net++, PSPNet, FPN, and DeepLabV3). In the case of obtaining similar segmentation results, with a smaller amount of parameters compared with State Of The Art (U-Net3+, ACSNet, PraNet, and CCBANet). We also use the farmland data provided by Sichuan Agricultural University for testing, the Dice is 93.88%, which has good generalization performance.