CT-HiffNet: A contour-texture hierarchical feature fusion network for cropland field parcel extraction from high-resolution remote sensing images

Automatically extracting cropland field parcels from remote sensing images is crucial for developing smart agriculture. However, notable spatio-spectral differences captured by multiple remote sensing sensors at different times led to the uncertain contour and texture features among large-scale cropland field parcel, posing challenges for robust and high-precision extraction. To address these challenges, we proposed a contour-texture hierarchical feature fusion network (CT-HiffNet) for cropland field parcels extraction from high-resolution remote sensing images. The CT-HiffNet consists of three modules: a hybrid module integrating attention and guidance method to thoroughly learn the internal texture features as well as external contour features of cropland field parcels; a deep residual shrinkage block for feature encoding to effectively eliminate redundant information during the extraction tasks; and a hierarchical information fusion decoder to enhance contour-texture feature interactions at different scales and minimize information loss during feature restoration. The CT-HiffNet was evaluated across four distinct agricultural landscape regions in China using GaoFen-2 images, as well as in six other global regions using Sentinel-2 and Google Earth images. The results showed that CT-HiffNet achieved an overall accuracy exceeding 80% across China, thereby proving its effectiveness in extracting cropland field parcels. Global experiment demonstrated the strong generalization and transferability of CT-HiffNet, despite variations in sensor types and image resolution. In particularly, the contour–texture feature effectively enhanced the boundary recognition of cropland field parcels, contributing to the model adaptability to different acquirement times of remote sensing images. Meanwhile, determining an appropriate sample size is crucial for the performance of CT-HiffNet.