A systematic review of integrating remote sensing and machine learning for near real-time water quality monitoring applications

ABSTRACT Conceptual diagram of three input streams: satellite remote sensing, IoT sensors, and citizen science. These inputs are processed using data fusion and a machine learning layer to provide near real-time water quality outputs and alerts. In semiarid and tropical regions, freshwater ecosystems are facing an increasing pressure from agricultural run-off, urban effluents, hydrologic alterations, and climate variability; yet, effective monitoring is hindered by sparse, irregular observations and limited institutional capacity. A total of 44 peer-reviewed studies from 2005 to 2025 (based on PRISMA 2020 guidelines) were collected for this systematic review, and the thematic synthesis of the results focuses on remote sensing (RS), machine learning (ML), Internet of Things (IoT) sensors, big data (BD) platforms, and citizen science to create near real-time monitoring capabilities for surface water quality. All studies originated from Sub-Saharan Africa (47%), Asia (32%), South America (14%), and other parts of the world (7%). RS had the highest percentage of use as a technology at 38%, ML and predictive modeling had a 25% use rate, IoT was at 14%, BD platforms accounted for 11%, citizen science made up 7%, and empirical methods at the watershed level made up the remaining 5%. RS-based models obtained estimates for chlorophyll a and turbidity using the Sentinel-2 and Landsat 8 satellites with R2 values between 0.72 and 0.94, while ML models (specifically random forest and support vector machine) maintained consistency in achieving root mean square error values below 10%. IoT and citizen science have the capacity to collect valuable local data but have not been adequately incorporated into satellite analytics or local calibrations with ML models. The main challenges to integrating these data sources into a coherent monitoring framework include (1) the absence of consistent time series; (2) the low resolution of satellite images; (3) very few ML models can be transferred between study areas; and (4) there is no interoperable platform to integrate and analyze multiple sources of information that can be used for surface water quality monitoring. Transitioning to a modular RS–ML–IoT framework that includes standardized citizen science methodologies and local calibrations of the ML models would enable scalable, context-specific early warning systems across all data-scarce regions, such as the East African Rift Valley.