Exploring Machine Learning's Potential for Estimating High Resolution Daily Snow Depth in Western Himalaya Using Passive Microwave Remote Sensing Data Sets

Abstract

Snow depth (SD) exhibits high spatiotemporal heterogeneity in Western Himalaya (WH), and its knowledge is essential for applications related to water resources, disaster management, climate, etc. However, due to inclement weather and rugged topographical conditions, only a sparse network of SD monitoring stations exists in WH. Spaceborne passive microwave (PMW) remote sensing data sets provides valuable information about SD; however, only a limited PMW SD studies that cover subregions of WH are available. Different machine learning (ML) methods viz. support vector machine, random forest, and Extremely Randomized Trees (ERT) were tested for estimating SD. Based on our preliminary assessment of these ML approaches, the current study utilizes ERT approach to estimate daily SD over the entire WH region. The ERT SD model is developed using PMW brightness temperature data sets from Advanced Microwave Scanning Radiometer-2 (AMSR-2), snow cover duration (SCD), and other auxiliary parameters (i.e., location, elevation, vegetation cover, etc.) during the winter period between 2012–2013 and 2019–2020. The data between 2012–2013 and 2017–2018 is used for training the model, whereas the data between 2018–2019 and 2019–2020 is used for testing the model. The results demonstrate: (a) The ERT SD model has shown improved SD estimates compared to the available PMW remote sensing-based operational SD products and empirical PMW SD models. (b) In general, with an increase in SD, the mean absolute error of SD retrievals has increased in all SD products/models. (c) Unlike the operational AMSR2 SD product, and Northern Hemisphere Machine Learning SD product, the ERT SD model retrievals have shown better consistency with MODIS snow cover. (d) The developed model has shown a wider range in SD retrievals as compared to other products considered in this study.