USDA耕作地図およびSentinel-2を学習データとしたPALSAR-2 ScanSARデータの機械学習による広域の水稲作付面積推定手法の開発

抄録

Early and accurate knowledge of crop acreage and yields is important for decision-making in food security policy and the grain-trading business. Crop acreage estimations achieved by applying machine learning to satellite data require the appropriate training data. Herein, we developed a new training data set for rice cultivation in California (USA) by combining the historical Cropland Data Layer provided by the U.S. Department of Agriculture (USDA) with the latest Sentinel-2 satellite data, which can be applied even when cropping patterns differ significantly from the past cropping patterns due to extreme events such as large-scale droughts. We then applied these training data to random forests with the Advanced Land Observing Satellite-2 (ALOS-2) data for classifying rice cultivation over a wide area with high accuracy, without the effects of clouds. An assessment of the proposed method’s accuracy was then conducted. The results demonstrated that our proposed method can estimate paddy rice acreage with <1 % error compared to the USDA statistics even in 2021, when a large-scale drought occurred. We also evaluated the relationship between the lead time to harvest and the accuracy of the proposed method’s area estimation; the results confirmed that the method estimated the area with an approx. 1 % estimation error even in late July, which was >1 month before the harvest in 2021. Our proposed method can thus be used for early and accurate estimations of paddy rice acreage even in years of drought extremes, and the method can be expected to be applied to food security policy, the grain-trading business, and more.