HYDROLOGICAL MODELLING OF A RIVER BASIN INTEGRATING REMOTE SENSING RETRIEVED WATER FLUXES

Abstract

Hydrological modelling is a complex process involving the simulation of various components or processes of the hydrological cycle. It faces several challenges due to the complexity and variability of these processes and the difficulty of continuous monitoring, particularly in developing countries with limited resources. Advances in technology, sensor networks, and modelling techniques, along with a comprehensive understanding of hydrological processes in diverse conditions, are essential to address these challenges. Remote sensing has become a critical tool in this field, offering valuable data from satellite or aerial observations. This technology helps estimate water fluxes, although it also introduces potential errors and biases that need evaluation for accuracy and reliability in hydrological modelling. Since in-situ measurements have limitations, alternative approaches using physics-based or data-driven hydrological models can offer a more comprehensive assessment of the performance of Satellite-based Precipitation Estimates (SPEs), thereby addressing the misalignments in spatial and temporal resolutions between SPEs and the hydrological processes.Moreover, efforts to integrate satellite-based soil moisture (SM) and evapotranspiration (ET) into hydrological model calibration need a thorough investigation. Recently, there has been an increase in the use of machine learning-based hydrological models, which can be useful for accurate streamflow simulations with less cost and time. To this end, this research aims to comprehensively assess the potential of satellite-retrieved water fluxes for hydrological modelling. This involves investigating the performance of SPEs over India for rainfall measurement and their error characteristics, and hydrological utilities. Further, both physics-based and machine-learning hydrological models should be improved by utilizing the remotely sensed products. Considering the limitations of ground-based datasets in capturing spatiotemporally continuous precipitation data, the performance of nine SPEs over 14 agro-climatic (ACL) zones in India is comprehensively evaluated against a reference gridded (observation) precipitation product developed by the India Meteorological Department (IMD). These nine SPEs were divided into two categories, i.e., Satellite-Only Precipitation Estimates (SOPEs), including CMORPH, PERSIANN, SM2RAIN-ASCAT, SM2RAIN-CCI, and SM2RAIN-GPM; and Gauge-Corrected Satellite Precipitation Estimates (GCSPEs), including CHIRPS, GSMAP, IMERG, and MSWEP. Various qualitative and quantitative statistical metrics were computed, whereas a multi-criteria decision-making (MCDM) approach was used to rank the datasets based on multiple performance indicators. A significant deviation in mean daily precipitation was observed in SPEs from reference data, particularly in zones with heavy rainfall. SPEs generally underestimated heavy rainfall events and overestimated low rainfall events. SM2RAIN-based products were more effective at detecting rainfall events, with SM2RAIN-GPM being the best among SOPEs. However, these products also had a high probability of falsely detecting rainfall events. Among GCSPEs, GSMaP was the best performer for accurately detecting rainfall events. Overall, the GPM-based products showed better agreement with reference data. The MCDM approach identified SM2RAIN-GPM as the superior performer among SOPEs and GSMaP among GCSPEs, with IMERG also performing well among GCSPEs. The analysis highlighted the inconsistency in performance across different statistical indicators, underscoring the complexity of evaluating precipitation estimates, which can be circumvented using the MCDM approach. In addition to quantifying bias and error in SPEs in comparison with the ground-based observations, it is crucial to further characterize these errors and identify their sources in the precipitation retrieval process, which can provide comprehensive information. To this end, this research employed error decomposition methods to disaggregate total bias in SPEs into components like over-hit bias, under-hit bias, miss bias, and false bias, alongside decomposing mean squared error into random and systematic parts. Utilizing datasets for topographic features (altitude and topographic complexity) and vegetation features (condition and dynamics), their influence on error statistics of SPEs was investigated for the nine SPEs. A higher random error was found in SPEs across most Indian land areas except for the western Himalayan region, suggesting potential improvement areas by correcting systematic biases. Higher random error percentages were observed for CMORPH and PERSIANN compared to SM2RAIN-based SOPEs, whereas GCSPEs exhibited high random error percentages across all ACL zones. The decomposition of total bias in different components reveals that the hit bias is the major contributor to the total bias. The maximum magnitude of missed bias is found in the CHIRPS and CMORPH products, whereas the highest amount of false bias is observed in CHIRPS, PERSIANN, and SM2RAIN-based products. The study also revealed that the performance of SPEs decreases with an increase in altitude and topographic complexity. Additionally, vegetation conditions positively correlate with SPEs' rainfall detection capability, whereas miss bias shows significant negative correlations with vegetation characteristics. These findings highlight the importance of considering topographic and vegetative factors in improving SPEs' performance and accuracy. Assessing the hydrological utility of SPEs remains challenging due to various sources of error associated with remote sensing measurements and atmospheric processes. Six SPEs (i.e., CHIRPS, CMORPH, GSMaP, IMERG, MSWEP, and PERSIANN), alongside IMD gridded precipitation product, were evaluated through a physics-based distributed Variable Infiltration Capacity (VIC) model over the Ashti catchment of Godavari Basin, India, for their performance in simulating the water balance components such as streamflow, SM, and ET. The VIC model was calibrated for 2002-2011 and validated for 2012-2017 against observed streamflow data under different calibration scenarios (daily, monthly, and high-flow conditions). The monthly simulations indicated that SPEs’ performances (KGE ranging from 0.84 to 0.88) were very close to or even better than gauge-based simulations (KGE = 0.85). However, the gauge-based precipitation product outperformed SPEs in simulating daily streamflow, although CMORPH and IMERG showed promising performances with KGE values of 0.74 and 0.77, respectively, in the overall period (2002-2017). CMORPH and IMERG also outperformed other SPEs in simulating high flows. In general, SPEs-forced VIC model overestimated low flows and underestimated high flows at daily timescale. IMERG showed the highest performance, closely followed by CMORPH, MSWEP, and GSMaP, while CHIRPS and PERSIANN were the least effective. For simulating ET using VIC model, CMORPH and IMERG were the best performers, whereas CHIRPS showed the poorest performance. For simulating SM, GSMaP, IMERG, and MSWEP were the best performers, whereas the SPEs viz., CHIRPS, CMORPH, and PERSIANN demonstrated relatively poor performance. The study highlights the variability in SPEs' performance across different hydrological components and conditions, underscoring the importance of selecting appropriate SPEs for specific hydrological applications. For calibrating distributed hydrological models, the simultaneous utilization of multiple satellite-retrieved products to constrain various hydrological state/flux variables is a rare practice. Recent advancements in remote sensing have provided valuable data for hydrological components like SM and ET, offering a unique opportunity for incorporating their spatial patterns and temporal dynamics within the calibration scheme. The study employed a multivariate calibration scheme to calibrate VIC model, comparing a streamflow-only calibration approach with a second approach integrating multiple remote sensing datasets of ET and SM alongside streamflow information. A bias-insensitive metric considering both spatial patterns and temporal dynamics of SM and ET was employed to assess the model performance. The results highlight the linkage of the performance of regionally calibrated hydrological models across local scales with the physical attributes of each region, emphasizing the need for careful consideration of land cover dynamics, topographical features, and local climatic variations in refining the model for accurate streamflow predictions. The model performance in streamflow prediction declines at the calibration site when calibrated with multivariate objective functions. However, a significant improvement in the model performance, for streamflow prediction, was observed at ungauged sites (sub-catchments with similar catchment characteristics to the catchment of the calibration site) when the model was calibrated with multivariate objective functions. Amongst multivariate calibration frameworks, the best model efficiencies were obtained for temporal dynamics-based multivariate strategy. Considering the cost and efforts associated with the physics-based modelling along with the unavailability of reliable observations over the developing countries, the potential of satellite-retrieved water fluxes through machine learning-based hydrological models should be assessed. Therefore, a study was conducted in six catchments from large and medium river basins in India through a Random Forest (RF)-based model, exploring how the incorporation of satellite-based SM and ET can improve discharge prediction. The data used included surface soil moisture product developed from backscatter coefficients measured by the ASCAT instrument, daily satellite-based basin-averaged precipitation data from IMERG V06, MODIS-based ET product, ground-based temperature, and river discharge from GloFAS, alongside the observed discharge for these basins. The importance of basin-averaged SM or Basin Water Index (BWI) in streamflow prediction was assessed through RF-based BWI-runoff, classical rainfall-runoff, and BWI-rainfall-runoff models. The results demonstrated a significant influence of SM in runoff generation, with a potential logarithmic relationship between BWI and discharge. The BWI-rainfall-runoff model performed the best, in general, indicating that satellite-based SM and rainfall data can complement each other effectively in hydrological modelling. In comparison with GloFAS discharge, the locally-developed models generally performed better, suggesting that global models might not be suitable for all regions and should be carefully evaluated before use in regional applications. The study underscores the effectiveness of integrating satellite-derived soil moisture and rainfall data in hydrological models for improved streamflow prediction, particularly in ungauged catchments in developing countries like India.