Optical-sar dual-branch time-series fusion rice recognition method and device based on cultivated land patch constraint

Abstract

The application provides a kind of optical-SAR dual-branch time sequence fusion rice identification method and device based on cultivated land patch constraint, the method comprises: generating high-precision cultivated land patch mask;Extract multi-temporal SAR time sequence characteristics and single-time optical characteristics;Sliding window slicing and filtering under spatial constraint are carried out using cultivated land patch mask;The dual-branch time sequence fusion network model containing parallel SAR encoding branch and optical encoding branch, hierarchical cross-modal feature fusion module and decoder is constructed and trained;The image to be identified is input into the model to output pixel-level probability map, and after binarization, cultivated land patch mask spatial cropping and majority voting in patch, the cultivated land patch level rice identification result is output.The application effectively suppresses pixel fragmentation and misclassification through full-process spatial constraint of cultivated land patch and asymmetric dual-modal time sequence fusion, and is suitable for rice planting area extraction in cloud and rain frequent and water network dense areas.