Machine learning-based mapping of fraction snow cover in High Mountain Asia by MODIS
High Mountain Asia(HMA)is the richest high altitude region in the world except for the poles in terms of glacier and snow resources,The accurate monitoring of HMA snowpack distribution is important for HMA snowmelt runoff simulation,climate change prediction and ecosystem evolution.Fractional Snow Cover(FSC)can quantitatively describe the extent of snow cover at the sub-image scale,and is more suitable for reflecting the distribution of snow in complex mountainous areas than binary snow.The objective of this study is to develop a new HMA snow area ratio inversion algorithm and integrate the algorithm into Google Earth Engine to prepare a set of long time series HMA snow area ratio products.Considering the influence of HMA topography and sub-bedding type on the accuracy of snow accumulation information extraction,this paper proposes a Multivariate Adaptive Regression Splines(MARS)model LC-MARS to invert the proportion of snow accumulation area in Asia by integrating topography correction and subland class feature extraction.The FSC extracted by Landsat 8 is used as the true value,and the LC-MARS model is tested for inversion FSC accuracy using binary and error validation methods,and the performance of linear regression models trained with the same training samples and the LC-MARS model for inversion HMAFSC accuracy is compared,and the accuracy of the FSC inversion of the LC-MARS model with SnowCCI and MOD10A1 is also compared.(1)The overall accuracy of FSC binary validation of LC-MARS model inversion showed that Accuracy and Recall were 93.4％and 97.1％,respectively,and the overall accuracy of error validation showed that RMSE was 0.148 and MAE was 0.093,both binary validation and error validation indicated that the FSC accuracy of LC-MARS model inversion was higher.(2)The LC-MARS model trained based on the same training samples has higher FSC accuracy than the linear regression model in forest area,vegetation and bare land inversions,indicating that the LC-MARS model is more suitable for FSC inversions in mountain and forest areas.(3)The overall RMSE of MOD10A1 is 0.178 and MAE is 0.096;the overall RMSE of SnowCCI is 0.247 and MAE is 0.131.The accuracy of FSC prepared by LC-MARS is higher than that of MOD10A1 and SnowCCI,indicating that FSC inversion by LC-MARS has some application value.The LC-MARS model can fit high-dimensional nonlinear relationships and significantly improve the inversion accuracy of FSC in mountain and forest areas.The computational efficiency of the LC-MARS model based on Google Earth Engine is high,and it is suitable for preparing FSC products of large scale long time series.In this study,the day-by-day MODIS FSC products of HMA from 2000 to 2021 were prepared based on the LC-MARS model,which provides important data support for the study of climate change,hydrological and water resources in HMA.
万方数据
维普
