Super-Resolution Cropland Mapping by Spectral and Spatial Training Samples Simulation

Medium spatial resolution remote sensing images are widely used for cropland mapping. In areas where cropland is fragmented, however, the limitation of the spatial resolution may lead to inaccurate or even impossible mapping of small croplands. Super-resolution mapping is an effective method to address this issue by transforming coarse-resolution fraction images, derived from spectral unmixing, into fine-resolution land cover maps. In practical applications, a crucial obstacle of this approach is the difficulty in collecting training samples for spectral unmixing and super-resolution mapping. To address this problem, this article proposed a novel super-resolution cropland mapping approach by simulating spectral and spatial training samples. Specially, a mixture spectral simulation method was used to generate training samples for the regression unmixing model to estimate cropland fraction images. A multilevel feature fusion U-NET model was proposed for super-resolution cropland mapping and was trained with simulated training samples considering fraction errors. The proposed method was tested in the Jianghan Plain, China, by generating 2.5-m cropland maps from the 10-m Sentinel-2 images. The results show that the proposed method can accurately extract more smaller and linear land cover features, preserve the spatial structure of the boundaries, and achieve higher accuracy than other cropland mapping methods. This method overcomes the dependency on actual sample collection in traditional methods, better utilizes spectral and spatial features in remote sensing data, and reduces the impact of spectral unmixing errors on the final fine-resolution cropland maps.