Research On Cropland Extraction From High-Resolution Remote Sensing Images Based On Swin Transformer

Semantic segmentation techniques in deep learning are currently an important method for extracting cropland from high-resolution remote sensing images.However,existing methods for cropland extraction from high-resolution remote sensing images still face difficulties in extracting fine boundary features and separating confusing terrain features.Therefore,a cropland extraction method for high-resolution remote sensing images based on hierarchical boundary enhancement is proposed in this paper,with the Swin Transformer being used as the backbone network.Boundary information in the features is enhanced by designing different boundary enhancement modules for low-level and high-level features.Furthermore,A method based on dual-path feature fusion is proposed in this paper to improve the recognition accuracy of confusing terrain features by fusing global and local information.The main contributions of this paper are as follows:(1)In this paper,a high-resolution remote sensing image cropland extraction method based on hierarchical boundary enhancement(HBRNet)is proposed.Because the low-level features of swing transformer contain more details,we design the boundary and subject separation module to extract the boundary and subject features with more details.Concerning the problem that high-level features in Swin Transformer contain more semantic information but lose boundary details,we design a boundary detail enhancement module,which fuses the boundary features extracted from lowlevel features with high-level features.After feature fusion,the generated feature maps interact with the body features extracted from low-level features to exchange boundary and body feature information.In addition,a cross-detail module is designed to preserve more boundary details and semantic information before the fusion of high-level and low-level features.The results of the experiments demonstrate that the IoU of the cropland category in the Agricultural dataset using HBRNet is 84.59%,while the IoU of the cropland category in the DeepGlobe dataset is 90.37%.(2)A method for extracting cultivated land from high-resolution remote sensing images based on dual path feature fusion(DBRNet)is proposed.The method for fusing high-level and low-level features in HBRNet is a local linear fusion method.DBRNet designs a spatial feature fusion module and a channel feature fusion module to achieve feature fusion of global and local information.The output feature map of the spatial feature fusion module is used as the input feature map of the channel feature fusion module.The spatial feature fusion module includes a global feature extraction path and a local feature extraction path based on cross convolution.The features output by the module are used as the importatnt parameters for dynamically assigning weights when fusing high-level and low-level features.The channel feature fusion module includes a global feature extraction path and a local feature extraction path based on point-wise convolution,effectively improving the recognition accuracy of easily confused features.In this paper,experiments are conducted on the publicly available land cover dataset DeepGlobe and the cropland dataset Agriculture that we constructed.The experimental results demonstrate that DBRNet method achieves an IoU of 85.49%on the Agriculture dataset and an IoU of 91.22%on the DeepGlobe dataset for the cropland category,with better accuracy than classical semantic segmentation methods.The method proposed in this paper has been applied in the construction of high-precision farmland maps in intelligent rice-wheat rotation farms in Wuhu City,Anhui Province and intelligent maize-wheat-legume crop rotation farms in Bozhou City,Anhui Province with good performance.