Research on crop remote sensing image segmentation method integrating CNN and transformer

Currently, when performing pixel-level semantic segmentation of crop planting in high-resolution images, it is difficult for deep convolutional neural networks to simultaneously capture global features and local detailed features at multiple scales in space. This can lead to blurred boundary contours between different farmland plots, as well as lower integrity within similar farmland areas. In view of the above research content, a model that integrates Transformer and CNN is proposed to classify and identify crops in the study area, using drone remote sensing images from a competition as the data source. (1) The model adopts a multi-level skip-connected encoder-decoder network architecture. The encoding part of the model uses an improved MobileNetV2 as the front-end feature extractor to extract local detail features, and then inputs the extracted features into Vision Transformer for global feature capture and further processing, thereby capturing details while retaining global context information. (2) The decoding part adopts the design of UNet, extracting features from different levels of the encoding part and directly transferring them to the corresponding levels of the decoding part through skip connections to ensure the retention of detail information, and using upsampling layers to gradually restore the spatial resolution of the image. In an experiment on a public competition dataset, the experimental results show that the MIoU of the network proposed in this paper reaches 85.96%, the PA reaches 92.30%, and the Dice value reaches 0.922, which has the highest segmentation accuracy compared with the comparison network.