XGBoost算法在河北省GPM卫星降水数据降尺度中的应用

Application of XGBoost algorithm in downscaling GPM satellite precipitation data in Hebei Province

针对GPM卫星降水产品分辨率相对较低问题,本研究引入XGBoost(eXtreme Gradient Boosting)回归算法,旨在改善河北省GPM降水数据的分辨率水平和应用精度。基于地形变化、土地利用、NDVI应用、经纬度和地面降水数据构建XGBoost的回归模型,然后运用网格搜索法进行参数优化,进而对河北省2020年GPM数据进行降尺度,利用独立验证点进行精度评估。结果表明,降尺度后的GPM数据与地面降水量在年尺度上具有高一致性,其R2达到了0.88,MAE(Mean Absolute Error,平均绝对误差)和RMSE(Root Mean Square Error,均方根差)分别为62.07 mm和85.72 mm。相比原始GPM数据,降尺度后的GPM数值精度R2提升了4.55%,MAE和RMSE分别降低了70.23%、71.14%。XGBoost模型能准确拟合星地多源降水数据之间非线性规律,在低精度、低分辨率GPM降水数据降尺度处理方面具有广阔应用前景。

This study aims to improve the resolution and accuracy of GPM precipitation data in Hebei Province to better reflect the actual ground situation.Using terrain,land use,NDVI,and longitude and latitude as covariates,an XGBoost regression algorithm was constructed to downscale the original GPM data for 2020,and the accuracy was evaluated through independent validation points.The results showed that the downscaled data had a strong correlation with the observed data on an annual scale,with an R2 of 0.88,MAE of 62.07 mm,and RMSE of 85.72 mm.Compared to the original GPM data,the downscaled precipitation data had a 4.55% improvement in numerical accuracy(R2),and a 70.23% and 71.14% reduction in MAE and RMSE,respectively.The XGBoost model exhibited good training accuracy and generalization ability,and had high precision and reliability in downscaling rough-level GPM precipitation data.It can provide more accurate data support for research in related fields.