A Frame on Snow Depth Reconstruction Based on Machine Learning Technique

Snow depth (SD) and snow water equivalent (SWE) are significant parameters in climate and hydrologic models. Successful estimation of SD (SWE) can improve the accuracy of snowmelt-runoff predictions and the management of water supplies. Currently, passive microwave (PMW) remote sensing is the most efficient way to monitor SD on global and regional scales; however, there are many challenges for accurate SD estimation. In this study, a new spatial dynamic method is developed by introducing random forest (RF) model, AMSR-2 TB and other auxiliary data. The main objective of this work is to produce long term SD dataset with the dynamic method using the Special Sensor Microwave Imager (SSM/I) and Special Sensor Microwave Imager/Sounder (SSMI/S) which span from 1987 to present. Through evaluation and analysis, the RF method performs better than traditional linear-fitting model. However, it tends to overestimate SD in shallow snow cover areas. Now, a preliminary spatial dynamic method (pixel-based model) is developed. For further work, the evaluation will be conducted to assess the feasibility of SD reconstruction. Moreover, to address overestimation over shallow snow areas, the snow depletion curve (SDC) incorporating SD and fractional snow cover (FSC) is expected to improve SD retrievals.