Determining the meteorological disaster of maize and soybean by machine learning algorithms with Sentinel-2 satellite images

Meteorological disasters such as windstorm, waterlogging, drought and so on, are crucial factors affecting crop production and farmers’ income. Agricultural insurance is one of the important strategies to protect the interests of farmers, especially in developing countries such as China. However, the accurate identification and quantification of meteorological disasters in large scale are still difficult issues for the popularization and development of agricultural insurance. One possible solution is to combine the high-resolution remote sensing satellite images with machine learning algorithms. In this study, we conducted the measurements for the yield of soybean and maize and determined the damage degrees of about 2000 fields in 2021. The Sentinel-2 satellite images were also collected in the same or adjacent date as the field measurements. The clustering algorithm was applied to amplify the field measurements. After that, three machine learning algorithms named LightGBM, XGboost and RandomForest were used to relate the surface reflectance, crop types, disaster damage degrees, and crop yields of soybean and maize. The results indicated that the accuracy of the XGBoost algorithm is better than the LightGBM and RandomForest. In addition, the present method obtained higher accuracy for the maize than the soybean, which indicates that meteorological and image data during crop growth periods should also be added in the yield estimation process, and the differences between crop loss mechanisms of different crops should be studied in the future.