Downscaling grace missions groundwater storage estimates using a machine learning model approach for climate change monitoring in the upper Zambezi catchment
Loading...
Date
2024
Authors
Shilengwe, Christopher
Journal Title
Journal ISSN
Volume Title
Publisher
The University of Zambia
Abstract
The Upper Zambezi Catchment (UZC) of Sub-Saharan Africa, is among the areas most susceptible to the effects of climate change. Significant challenges for water management are brought on by rising temperatures, shifting rainfall patterns and an increase in the frequency of droughts. In this area, groundwater is an essential resource for ecosystems and human needs. Effective water
management, particularly in the context of climate change, depends on groundwater storage (GWS) monitoring. By measuring gravity anomalies, the GRACE and GRACE-FO satellite
missions have completely changed how terrestrial water storage (TWS) is estimated globally. However, the use of GWS data for local and regional water management is restricted because of
the coarse spatial resolution of these missions. The primary objective of this research was to use open-source remotely sensed data and machine learning (ML) to downscale the GRACE/GRACE-FO GWS estimates to a finer spatial resolution (5 km). Several hydrometeorological variables, including evapotranspiration (ET), land surface temperature (LST), normalised difference vegetation index (NDVI), enhanced vegetation index (EVI), soil moisture and rainfall, were integrated during the downscaling process using the Extreme Gradient Boosting (XGBoost) and Random Forest (RF) algorithms. This method was unique in that it used a time-series split for cross-validation and Z-score normalisation. This made it possible to evaluate the model's performance with confidence. The findings showed that RF regularly performed better than XGBoost in downscaling GWS estimates, yielding more precise predictions because of its capacity to control intricate correlations between variables and prevent overfitting. The correlation coefficients between the coarse input data and the downscaled data in the Barotse sub-catchment for Nash-Sutcliffe efficiency (NSE) were 0.81 (XGBoost) and 0.89 (RF). Segmenting the UZC into smaller sub-catchments improved model accuracy by enabling the analysis to focus on smaller regions with more homogeneous hydrogeological characteristics. In regions with unconfined aquifers, the downscaling greatly enhanced the GWS estimates. The model, however had challenges in areas with confined aquifers, which usually have low hydraulic connectivity and slow recharge rates. One of the study's main findings was that groundwater recharge is greatly decreased below a
rainfall threshold of 614.39 millimetres during the wet season. This threshold is essential for anticipating periods of low recharge and planning for water scarcity. The study also showed the
extent to which the downscaled GRACE/GRACE-FO data captured drought conditions in groundwater, especially during El Niño years (2015, 2016, 2018, 2019 and 2023), underscoring themissions' potential for drought monitoring. The study found that, on average, GWS decreased throughout the UZC between 2009 and 2023, with anomalies varying from -433 millimetres to
+264 millimetres. This research underscores the potential of machine learning in enhancing the spatial resolution and accuracy of GWS estimates, making GRACE/GRACE-FO data more useful for regional water management. These findings have significant implications for managing groundwater resources in the UZC and similar regions experiencing climate variability and increasing water demand.
Description
Thesis of Master of Science in Integrated Water Resource Management