Remote Sensing Data Assimilation to Improve the Seasonal Snow Cover Simulations Over the Heihe River Basin, Northwest China

Abstract

The reliability of seasonal snow cover information is constrained by limitation of in situ observations and uncertainties in remote sensing data and model simulations in alpine region, thus posing important challenges to understanding the climate system and water resource management in alpine region. Here, the assimilation of daily cloud-free Moderate Resolution Imaging Spectroradiometer (MODIS) normalised difference snow index (NDSI) product into an intermediate complexity snow mass and energy balance model—Flexible Snow Model (version FSM2_MO)—was implemented. The aim is to improve the model simulations of seasonal snow cover (snow-covered extent; SCE, snow depth; SD, snow water equivalent; SWE, and snowmelt runoff; SMR) in the alpine region (a case of the upper-middle reaches of the Heihe River basin, Northwest China). The results indicate comprehensive improvement in the simulation of SCE, SD, and SMR in the study area through data assimilation, with the ability to significantly reduce prior biases of the FSM2_MO. Based on the independent daily cloud-free Advanced Very High Resolution Radiometer (AVHRR) SCE product, the updated SCE simulation (i.e., data assimilation) showed a reduction in mean absolute error (MAE) from 10.46% to 7.16%, root mean square error (RMSE) from 16.14% to 12.26%, and an increase in Pearson's correlation coefficient (CC) from 0.18 to 0.67 compared with the open loop simulation (i.e., without assimilation). The evaluation results of SD observation data showed that data assimilation improved SD simulation compared with the open loop run (OL). And utilising the monthly discharge observations at the Yingluoxia hydrological station, data assimilation slightly improved the SMR simulation. The updated SMR simulation achieved a CC of 0.91, Nash-Sutcliffe efficiency coefficient (NSE) of 0.73, and Kling-Gupta efficiency coefficient (KGE) of 0.76. Moreover, the Landsat 8-derived snow cover map and Sentinel-1-derived SD also indicated that the updated simulation effectively filled in the missing snow cover and removed the superfluous snow cover predicted by the OL simulation in terms of spatial distribution.