Cross-Scale Feature Selection Transformer for Street-Level Crop Classification

This study introduces a Cross-Scale Feature Selection Transformer (CFST) to enhance crop classification accuracy using street view (SV) images. Firstly, semantic segmentation techniques are employed to overcome challenges posed by the sky background. CFST leverages the hierarchical feature extraction capability of the CrossFormer architecture and incorporates a weight selection module. Experimental evaluation conducted on the Google Street View dataset showcases an impressive accuracy of 94.95% in crop classification with CFST, while significantly reducing model parameters. The comparative analysis further substantiates the superior performance of CFST, unveiling the untapped potential of SV images in the agricultural domain. This research makes a substantial contribution to the field and paves the way for the development of precise and lightweight crop classification models utilizing SV images.