Generating Fine-resolution Crop Maps Based on a Geographic Position-Modulated Super-Resolution Mapping Network

Abstract

Accurately mapping crop distribution is essential for sustainable agriculture, yet medium-to-low resolution remote sensing data, despite their long-term availability and low cost, are limited in fine-scale applications. Deep learning-based super-resolution mapping offers a promising solution by inferring sub-pixel crop distributions. However, conventional models either rely on spectral features or incorporate priors through shallow feature interactions, limiting their ability to capture geographically varying spectral and phenological characteristics in large-scale applications. To address this, we proposed a geographic position-modulated super-resolution crop mapping network (GPMNet), which exploits geographically stable coordinate priors as spatial constraints for adaptive feature learning. The proposed geographic position-modulated module converts these priors into affine transformation parameters that dynamically modulate spectral features, improving discrimination under geographically heterogeneous conditions. Tested in the three northeastern provinces of China for rice, maize, and soybean mapping, GPMNet outperformed comparison methods in visual and quantitative assessments. Notably, it reduces confusion between maize and soybean, yielding 4.84% (F1-score) and 6.45% improvements for soybean classification. Furthermore, GPMNet shows strong scalability and generalization, producing 10 m resolution crop maps for 2014 and 2015 that align well with statistical data and provide finer spatial detail. These results highlight its effectiveness for high-precision, large-scale crop mapping and historical fine-resolution reconstruction.