Synergistic approach for streamflow forecasting in a glacierized catchment of western Himalaya using earth observation and machine learning techniques

Accurate forecast of daily streamflow of the major Himalayan rivers is crucial for understanding changes in their hydrological regimes caused by precipitation variability, as well as their potential implications for hydropower generation and flood management in the downstream region. However, the study pertaining to daily streamflow forecasting in the western Himalayas is lacking, primarily due to the limited availability and accessibility of long-term daily gauged data. In this study, four machine learning (ML) models, namely support vector machine, random forest, deep neural networks, and bi-directional long-short-term memory (BLSTM), were employed to forecast daily streamflow in the Sutlej river basin. The models were trained (11 years) and tested (2 years) using publicly accessible earth observation and reanalysis datasets for a period of 13 hydrological years, grouped into four distinct scenarios, denoted as M1, M2, M3, and M4. In each scenario, feature parameters were generated from the pool of six hydro-meteorological variables namely- precipitation (P), air temperature ( \({{\text{T}}}_{{\text{a}}}\) ${\text{T}}_{\text{a}}$ ), evapotranspiration (ET), snowfall, snow cover area (SCA), and discharge. The optimal feature combinations (e.g., scenarios) were identified using the time lag cross correlation technique. All of the ML models demonstrated reasonable performance (NSE > 0.82, R > 0.91, and RMSE < 101.2 m³/s) in simulating streamflow during testing period, the BLSTM model showed a slightly superior level of accuracy across all scenarios (NSE = 0.97, R = 0.99 and RMSE = 42.7 m³/s for M4). The BLSTM was further used to reconstruct daily timeseries discharge for the period 2012–2021. The variability in \({{\text{T}}}_{{\text{a}}}\) ${\text{T}}_{\text{a}}$ , ET, and P exerts a greater influence on streamflow generation in the glacierized catchment in comparison to snowfall. The observed phenomenon can be ascribed to the dampening effect of snowfall on discharge, as their effects become apparent within a timeframe of five to seven days.