Satellites and Machine Learning Reveal 25 Years of Hidden Water Quality Shifts in an Ethiopian Drinking-Water Reservoir

A reservoir that supplies drinking water to millions of people in Addis Ababa has undergone a quiet but profound transformation over the past quarter century, and much of that change would have gone unnoticed by conventional monitoring. A new study of the Gefersa Reservoir, published in Environmental Monitoring and Assessment, has reconstructed twenty-five years of water quality history using satellite imagery, historical field measurements, and ensemble machine learning, revealing that the seasonal rhythms that once governed the reservoir’s ecology are breaking down in ways that carry direct consequences for water treatment and public health.

The research, led by Belachew Hirpa Lemma of Addis Ababa Science and Technology University together with Israel Tessema Lewte and Fekadu Fufa Feyessa of Jimma University, focused on five so-called optically active water quality parameters: chlorophyll-a, total suspended solids, turbidity, colored dissolved organic matter, and Secchi disk depth. These are the properties of water that leave measurable fingerprints in the light reflected from its surface, which makes them uniquely suited to satellite retrieval. Chlorophyll-a signals algal biomass, suspended solids and turbidity track sediment loads, colored dissolved organic matter reflects decaying organic material washed in from the catchment, and Secchi disk depth measures how far light penetrates through the water column.

What makes the Gefersa Reservoir scientifically interesting, and operationally worrying, is that it is a sediment-dominated water body sitting in a subtropical highland catchment. Reservoirs of this type face intensifying pressures from catchment erosion, pollution, unsustainable land practices, and a shifting climate, yet long-term water quality records in such settings are almost nonexistent. Ground-based monitoring in data-sparse regions tends to be intermittent, expensive, and vulnerable to gaps in funding, which means that slow, multi-decadal degradation can proceed invisibly. The Ethiopian team set out to close exactly that gap by building a monitoring framework that could look backward in time across decades.

The methodological core of the study is a seasonally optimized, cross-sensor harmonized remote sensing pipeline built on the Google Earth Engine platform. The researchers integrated historical in situ measurements with meteorological records and satellite data spanning 2001 to 2025, harmonizing observations from different sensors so that the long record remained internally consistent despite changes in instrumentation over the years. On top of that optical foundation they trained ensemble machine learning models, combining algorithms of the kind pioneered in random forests and gradient boosting, to translate spectral reflectance into estimates of each water quality parameter. The ensemble approach proved highly effective, achieving coefficients of determination of up to 0.878, a level of predictive performance that gives real confidence in the reconstructed time series.

The twenty-five year record revealed patterns that would have been impossible to detect from scattered field campaigns. For most of the study period, the reservoir behaved as expected for a monsoon-influenced highland system: concentrations of total suspended solids, turbidity, and colored dissolved organic matter peaked during the wet season, when rains stripped sediment and organic material from the catchment and flushed it into the reservoir. That classic wet-season dominance of sediment-related parameters is the pattern water managers in the region have long planned around, concentrating erosion control efforts and treatment capacity on the months of heaviest runoff.

But the new analysis shows that this predictable seasonal structure is eroding. The differences between wet and dry season conditions diminished markedly over the study period, driven largely by dramatic increases in dry season chlorophyll-a and colored dissolved organic matter, which rose by 147 to 173 percent. In other words, the dry season, once the reservoir’s period of relative recovery and clarity, has become nearly as biologically and chemically active as the rainy months. Algal growth and organic matter loading are no longer confined to the season of nutrient influx; they now persist through the dry months as well.

Equally striking is a temporal shift in when chlorophyll-a reaches its peak. Before 2015, algal biomass in the reservoir was dominated by dry season blooms; after that year, chlorophyll-a shifted to wet season dominance. At the same time, the researchers found that chlorophyll-a became significantly decoupled from seasonal climate cycles, meaning that algal dynamics in the reservoir are no longer simply tracking rainfall and temperature as they once did. Such decoupling is a hallmark of ecosystems pushed past a threshold, where internal feedbacks, accumulated nutrients in sediments, and altered catchment dynamics begin to override the external climatic drivers that previously organized the system.

The trophic state analysis drives home how serious these changes are. Using the Trophic State Index, a standard limnological metric introduced by Carlson in 1977 that classifies water bodies by their nutrient and algal status, the team found that the reservoir has remained persistently eutrophic to hypereutrophic throughout the study period, meaning it is chronically overloaded with nutrients and prone to dense algal growth. More alarmingly, dry season trophic state increased by 7.2 index units over the twenty-five years and, after 2011, began exceeding wet season values. A drinking water source that grows more eutrophic during its dry season, precisely when dilution capacity is lowest and treatment demands are highest, represents a compounding operational challenge for the utilities responsible for delivering safe water to Addis Ababa.

The implications extend well beyond a single reservoir on the outskirts of the Ethiopian capital. Eutrophic and hypereutrophic conditions elevate the risk of harmful algal blooms, increase the organic matter load that water treatment plants must remove, raise chemical treatment costs, and can promote the formation of disinfection byproducts when organic-rich water is chlorinated. Sediment-dominated systems like Gefersa face the additional burden of high turbidity, which interferes with treatment processes and reduces the light penetration that healthy aquatic ecosystems depend on. The study’s finding that wet and dry season problems are converging suggests that treatment plants and catchment managers can no longer schedule their responses around a predictable seasonal calendar.

The authors argue that their results make the case for seasonally adaptive monitoring: surveillance strategies that adjust to the evolving patterns rather than assuming static seasonal norms. Targeted erosion control in the catchment, water treatment protocols tuned to the new timing of algal and sediment pulses, and continuous satellite-based tracking all follow naturally from the framework demonstrated here. Because the approach relies on freely available satellite data, cloud computing, and machine learning models that can be retrained as new field measurements arrive, it offers a realistic template for other data-sparse reservoirs across the subtropical highlands, where drinking water security increasingly depends on seeing changes that no one is standing at the water’s edge to measure.

Subject of Research: Long-term remote sensing of optically active water quality parameters and eutrophication dynamics in a subtropical highland drinking-water reservoir in Ethiopia

Article Title: Seasonal and temporal dynamics of optically active water quality parameters in Gefersa Reservoir, Ethiopia

Article References: Lemma, B. H., Lewte, I. T., & Feyessa, F. F. (2026). Seasonal and temporal dynamics of optically active water quality parameters in Gefersa Reservoir, Ethiopia. Environmental Monitoring and Assessment, 198(10), Article 1126. https://doi.org/10.1007/s10661-026-15893-y

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15893-y

Keywords: Gefersa Reservoir, water quality, remote sensing, machine learning, chlorophyll-a, eutrophication, Trophic State Index, seasonal dynamics, Google Earth Engine, drinking water, Ethiopia, turbidity