A Sample Augmentation Method for Large-scale Rice Mapping Using Sentinel-1 and Sentinel-2 on the Google Earth Engine

Accurate and timely rice mapping is essential for agricultural decision-making and yield estimation for this key staple crop. However, the major challenge in large-scale rice mapping is constrained by the limited availability of labeled training samples and regional variability in the thresholds used for automatic sample generation. To address these challenges, in this study, AugSamp-RF, a sample augmentation method that integrates time-series Sentinel-1 and Sentinel-2 data for large-scale rice mapping with a random forest classifier, was developed. The methodological innovation focuses on adaptive extraction of optimal thresholds based on spatial histograms of spectral similarity values (SSVs), enabling high-confidence sample augmentation with minimal initial samples. Application of this adaptive thresholding procedure revealed clear regional differences in the optimal SSV thresholds. Specifically, the thresholds ranged from 0.11 to 0.29 in Heilongjiang but were approximately 0.50 in Guangdong and Hunan. This contrast indicates that lower thresholds are more suitable for homogeneous agricultural landscapes, whereas higher thresholds are required in heterogeneous and fragmented regions. The quality of the augmented samples was evaluated using the corresponding classification results, which achieved overall accuracies of 91.97% in the relatively homogeneous agricultural landscape of Heilongjiang and 92.76% and 91.41% in the more fragmented landscapes of Guangdong and Hunan, respectively. These results demonstrate the consistent effectiveness of the adaptive thresholding approach for sample augmentation across agricultural landscapes with different degrees of spatial heterogeneity. Furthermore, a feature-combination experiment in Heilongjiang showed that integrating vegetation indices (VIs) from key phenological stages with time-series vertical transmit- horizontal receive (VH) backscatter achieved an overall accuracy of 96.42%, only 0.08% lower than that obtained using full growing-season VIs, while reducing the processing time for an area of approximately 14 520 km² from 2 min 18 s to 17 s. The best-performing feature combinations achieved overall accuracies of 96.50% in Heilongjiang, 96.33% in Guangdong, and 93.05% in Hunan. AugSamp-RF also achieved higher accuracies than SVM and KNN in all three study areas and outperformed the evaluated rice products in Heilongjiang and Guangdong. Overall, AugSamp-RF is effective and applicable to both homogeneous and fragmented rice-growing regions, including cloud-prone areas, provided that the optimal SSV threshold is adaptively determined using representative regional data.