近年来全球变暖导致强对流天气日益加剧，冰雹灾害已成为农业生产的主要灾害之一。开展棉花冰雹灾情遥感评估对防灾减损、保险理赔、种植结构调整均具有重要意义。以2019年8月23日新疆准噶尔盆地西南部的奎屯河流域的棉花雹灾为研究对象，基于野外实测样本和雹灾前后多时相Sentinel-2遥感影像数据，分析雹灾前后的多种植被指数的动态变化规律，筛选能有效表征雹灾灾情的敏感植被指数差值特征组合，利用逻辑回归、决策树、梯度提升决策树、随机森林4种机器学习算法自动提取棉花雹灾的受灾范围与灾情等级，并利用野外实测样本进行精度对比分析。结果表明：单一植被指数中NDVI对雹灾的指示效果最佳，总体精度为84.39%，Kappa系数为0.75；多时相植被指数差值组合对雹灾的指示性显著优于单一植被指数；结合雹灾前后的植被指数差值时序特征，8月30日与8月20日差值对雹灾的指示性明显强于8月25日与8月20日的差值，说明雹灾灾情等级遥感监测有必要考虑灾后棉花植株的自我恢复能力，待灾情稳定后监测为宜；利用灾前灾后多种植被指数差值组合和随机森林分类算法的棉花雹灾灾情等级监测效果最佳，总体精度达到了89.51%，Kappa系数为0.83。基于多时相Sentinel-2影像能有效评估棉花雹灾的受灾范围以及灾情程度。

In recent years， global warming has led to an increase in strong convective weather， and hail disaster has become one of the main disasters in agricultural production. Carrying out remote sensing assessment of cotton hail disaster is of great significance for disaster prevention and mitigation， insurance claims and planting structure adjustment. Taking the cotton hail disaster in the Kuitun River Basin in the southwest of Junggar Basin， Xinjiang， on 23 August， 2019 as the research object， with the support of field measured samples， the multi-temporal Sentinel-2 remote sensing images before and after the hail disaster were obtained. We analyzed the dynamic changes of various vegetation indexes before and after the hail disaster， and screened the sensitive vegetation index difference feature combinations which can effectively characterize the hail disaster. The range and grade of cotton hail disaster were automatically extracted by using machine learning algorithms such as logical regression， decision tree， gradient lifting decision tree and random forest， and the accuracy was compared and analyzed via field measured samples. The results showed that NDVI was the best indicator of hail disaster among single vegetation index， with an overall accuracy of 84.39% and a Kappa coefficient of 0.75. The combination of multi-temporal vegetation index differences was significantly more indicative for hail disaster than that of single vegetation index. Compared with the time series characteristics of vegetation index differences before and after hail disaster， the indicative of hail disaster between August 30 and August 20 was obviously stronger than that between August 25 and August 20， which indicated that it was necessary to consider the self-recovery ability of cotton plants after hail disaster grade for remote sensing monitoring. The combination of the pre- and post-disaster vegetation indices and the random forest classification algorithm was the most effective methods in monitoring the level of cotton hail disaster level， with an overall accuracy of 89.51% and a Kappa coefficient of 0.83. In conclusion， the extent and degree of cotton hail disaster can be effectively evaluated based on multi-temporal Sentinel-2 image.