基于FY-3D反演土壤湿度的机器学习模型对比OA

Accurate monitoring of soil moisture is essential for assessing persistent drought conditions in Northeast China.To investigate the applicability and limitations of machine learning models for drought monitor-ing under different underlying surface conditions,FY-3D remote sensing data are integrated in this study with in situ soil moisture observations.Factors such as land cover type,soil depth,and crop growth stages are consid-ered to select relevant drought indices.Four machine learning algorithms,Radial Basis Function(RBF),eX-treme Gradient Boosting(XGB),Light Gradient Boosting Machine(LightGBM),and Random Forest(RF)are employed to construct and comparatively analyze soil moisture retrieval models.The results indicate that:①the RF model demonstrates significantly superior stability and accuracy compared to the other models,with an R² ranging from 0.64 to 0.81,exceeding those of LightGBM,XGB,and RBF by 0.14-0.35;②model perfor-mance varies with soil depth,with the 0-20 cm layer achieving the best performance(average RMSE of 9.52%),outperforming the 10 cm(10.19%)and 20 cm(9.79%)layers,and achieving an average accuracy of 88.03%,which is 1.95%higher than that at 10 cm;③across different land cover types(forest,cropland,and grassland),the variation pattern of model accuracy over crop growth stages is highly consistent.Higher accura-cy is observed during frozen soil and bare soil periods(>88.70%),with the highest value occurring in forest areas during the frozen soil period(94.30%).Accuracy is slightly lower during sowing and maturity stages(80.90%-84.10%),and lowest during the jointing stage,particularly in grassland areas(79.70%).