Automated in-season mapping of winter wheat in China with training data generation and model transfer

Accurate and timely information on winter wheat spatial distribution is essential for food security and environmental sustainability. However, high-quality nation-wide winter wheat products at high resolutions are still scarce around the world, and the approaches for winter wheat mapping are generally constrained by the lack of sufficient and representative training data. In this study, a knowledge-based approach based on spectral and polarization information from critical stages of winter wheat, was proposed to extract high-quality training data of winter wheat, thereby supporting winter wheat mapping with machine learning classifiers. Additionally, classification model trained by the generated training data was transferred across years to achieve the in-season mapping of winter wheat. Two-year classification scenarios based on the automated training data generation (ATDG) or model transfer (MT) were designed to evaluate the quality of automatically generated training data, the performance of model transfer, the contribution of optical and radar data, and the earliest timing for winter wheat mapping over China. With the ATDG and MT, the first 10-m resolution maps of winter wheat over China (ChinaWheat10) were produced for three consecutive years (2020 & 2021 by ATDG; 2021 & 2022 by MT). For ATDG and MT, the combined features of Sentinel-1 and Sentinel-2 yielded the highest overall accuracies with the random forest classifier. Specifically, winter wheat mapping with the ATDG achieved the highest F1-score of 0.94 for both 2020 and 2021. The MT reached a comparable F1-score of 0.94 and 0.93 for 2021 and 2022, and winter wheat maps with the F1-score of 0.93 and 0.92 could be produced as early as April (two months ahead of harvesting). Besides well-delineated winter wheat parcels, the estimated areas of ChinaWheat10 aligned well with the agricultural census data at the provincial (R2 ≥ 0.95) and municipal (R2 ≥ 0.91) levels. These findings suggest the proposed approaches have a great potential for accurate, cost-effective and high-resolution in-season mapping of winter wheat over large regions.