Combination Manner of Sampling Method and Model Structure: The Key Factor for Rice Mapping Based on Sentinel-1 Images Using Data-Driven Machine Learning

Agricultural remote sensing community is increasingly focusing on enhancing crop mapping accuracy by improving data-driven machine-learning model structures, yet ignoring impact of sampling–model structure combination on it, which may prevent full utilization of input data, especially for synthetic aperture radar images with fewer crop prior features. Consequently, this article took rice as target crop, and systematically performed rice mapping experiments based on Sentinel-1 images to assess mapping accuracies, model learning results, and model uncertainty under different sampling–model structures combinations. The sampling methods included pixel sampling in buffer or nonbuffer mode with equal proportion and equal quantity (pixel sample), as well as panoramic information sampling (image sample). The included model structures mainly focused on the models commonly used in rice mapping [i.e., Random Forest (RF) and Unet as traditional pixel and image data-driven machine-learning models], and related advanced model structures (i.e., popular transformer and Unet's variant, TransUnet, served as advanced model structures compared to the corresponding model structures commonly used in rice mapping). The experimental results showed that, when image sample was annotated well, both Unet and TransUnet were more suitable for rice mapping based on Sentinel-1 images, and their overall accuracies could reach 95% as sample size increased. Otherwise, when pixel sample size exceeded 100 000-level, nonbuffer equal proportion sampling–advanced transformer combination could be the currently optimal selection over the combination of this sampling method and RF, and its overall accuracy could reach 91% as sample size increased. Besides, it was worth noting that for data-driven machine-learning models commonly used in rice mapping, key factors for pixel data-driven ones to improve mapping accuracy was model structure upgrade, while for image data-driven ones, richness of image samples was more important.