Satellite Rainfall Maps Tested Against 95 Ethiopian Gauges Reveal Clear Winners

Ethiopia’s farms, rivers and ecosystems live or die by the timing and amount of rainfall, yet the country’s ground-based weather stations are too sparse to track it reliably. A new evaluation published in Theoretical and Applied Climatology has put six widely used satellite precipitation products through one of the most thorough tests yet conducted over the Horn of Africa, comparing their estimates against daily rain gauge records from 95 Ethiopian stations between 2001 and 2024. The verdict is nuanced but practical: no single product wins everywhere, but CHIRPS version 3, CHIRPS version 2 and IMERG consistently rise to the top, while GSMaP systematically underestimates rainfall and the reanalysis-based AgERA5 overestimates it.

The stakes could hardly be higher. More than 80 percent of Ethiopia’s population depends on rain-fed agriculture, and the country’s history is scarred by rainfall extremes, from the devastating 1983–1984 famine to the 2015 El Niño-induced drought that crippled harvests. Ethiopia’s climate is shaped by the Intertropical Convergence Zone, the Indian Ocean Dipole and El Niño-Southern Oscillation, layered on top of elevation gradients that run from the Danakil Depression below sea level to the Simien Mountains above 4,000 metres. Rainfall ranges from under 300 millimetres per year in the Afar and Somali lowlands to more than 2,000 millimetres in the southwestern highlands, and falls in three distinct seasons: the main Kiremt rains from June to September, the shorter Belg rains from March to May, and the dry Bega season from October to February.

Satellite precipitation products promise to fill the gaps left by the gauge network, which is densest in the central and northern highlands and thinnest precisely where monitoring is hardest. The research team, led by Getachew Mehabie Mulualem, evaluated CHIRPS_v2, CHIRPS_v3, IMERG, GSMaP, PERSIANN and AgERA5, a dataset derived from the ERA5 atmospheric reanalysis. These products differ fundamentally in how they work: some rely on infrared measurements of cold cloud tops, others on passive microwave sensors of the kind flown on the Global Precipitation Measurement mission constellation, and some blend satellite retrievals with gauge-based corrections. AgERA5, by contrast, comes from a numerical weather prediction framework rather than direct satellite retrieval.

The validation relied on daily observations from the Ethiopian Meteorological Institute, quality-controlled by removing extreme outliers beyond three standard deviations and discarding station-years with more than half of daily records missing. After screening, 1,633 site-years remained. The team matched each gauge to its corresponding satellite grid cell and stratified the analysis by elevation zone, using the Shuttle Radar Topography Mission to divide the country into lowlands below 1,500 metres, mid-elevations between 1,500 and 2,500 metres, and highlands above 2,500 metres, as well as by land cover classes from the European Space Agency’s WorldCover 2020 map. Of the 95 stations, 53 sit at mid-elevation, 30 in lowlands and 12 in highlands, with cropland the most common surrounding land cover.

The headline result concerns temporal scale. At the daily level, agreement between satellites and gauges was weak for every product, with coefficients of determination around 0.1, meaning the products explained only about a tenth of the day-to-day variance in what gauges recorded. Daily root-mean-square errors ranged from 12.0 to 16.3 millimetres, and relative biases were substantial, spanning from minus 36.2 percent for GSMaP, which badly underestimates, to plus 32.8 percent for CHIRPS_v2, which overestimates. When estimates were aggregated to monthly totals, however, correlations jumped dramatically to R² values between 0.62 and 0.75, and at annual scales they ranged from 0.28 to 0.60. The pattern reflects a well-known property of rainfall retrieval: errors in detecting or timing individual storms partly cancel out when data are accumulated, so monthly and seasonal totals are far more trustworthy than any single day’s satellite estimate.

Across the annual cycle, CHIRPS_v2, CHIRPS_v3 and IMERG proved the most reliable, combining high correlation, low bias and strong rainfall detection skill. CHIRPS_v2 and IMERG approached near-zero bias at the monthly scale, while CHIRPS_v3 showed a slight overestimation. Seasonal analysis sharpened the picture: all products performed best during the wet Kiremt season, when heavy rainfall provides a strong signal, and weakest during the dry Bega season, when small rainfall totals amplify the relative impact of detection errors. GSMaP’s underestimation was most pronounced in wet conditions, while AgERA5’s overestimation peaked during the wettest months, a limitation consistent with known problems of reanalysis precipitation in data-sparse, topographically complex regions.

Elevation and land cover both left clear fingerprints on product accuracy. Correlations generally declined with increasing elevation for most products, and every product showed somewhat lower probability of detection and higher false-alarm ratios in the highlands than in the lowlands. GSMaP’s underestimation grew worse with altitude, while AgERA5 overestimated rainfall in every zone, most severely at higher elevations. Land cover mattered too: false-alarm rates rose in densely vegetated areas such as tree cover, shrubland and grassland, where complex canopies complicate the remote-sensing signal, whereas built-up and cropland areas yielded better detection statistics. In bare, sparsely vegetated regions, all products showed stronger correlations, largely because rainfall there is low and less variable.

The study went beyond simple correlation by examining whether the products reproduce the dominant pattern of year-to-year rainfall variability across the whole station network. Using principal component analysis of standardized annual rainfall anomalies, the researchers extracted the leading mode of interannual variability from the gauges and compared it with each product’s version. CHIRPS_v2 and CHIRPS_v3 tracked the gauge-derived signal most faithfully, with correlations of 0.86 for both, followed by AgERA5 at 0.84, GSMaP at 0.77, IMERG at 0.769 and PERSIANN at 0.68. The products successfully reproduced major anomalies such as the pronounced wet phase around 2006 and the strong dry phase around 2009, though differences in amplitude and, in some years, even in the direction of departure emerged, particularly in the later part of the record. A 95 percent confidence interval around the gauge signal, built from 500 bootstrap resamples of the station network, quantified uncertainty in the network composition itself.

Distributional tests added a subtle but important caveat: matching the statistical shape of rainfall anomalies does not guarantee matching the amounts. GSMaP produced the lowest Kolmogorov–Smirnov statistics against the gauge anomaly distributions across daily, monthly and annual scales, indicating the closest overall distributional match, yet it simultaneously and persistently underestimated rainfall totals. At the daily scale, CHIRPS_v3 and PERSIANN showed the largest distributional differences, with KS statistics of roughly 0.27 to 0.34, while GSMaP and IMERG ranged from about 0.13 to 0.24. Aggregation again helped: at monthly and annual scales the KS statistics fell to approximately 0.07 to 0.13. The lesson, the authors stress, is that magnitude accuracy and distributional agreement are complementary measures that must be assessed separately.

The practical guidance that emerges is application-specific rather than one-size-fits-all. For long-term rainfall monitoring and drought-related work, CHIRPS_v3 appears the strongest choice, with CHIRPS_v2 remaining competitive in semi-arid conditions. IMERG stands out where detecting whether rain actually fell on a given day matters most, such as event-based applications, though its slight tendency to underestimate suggests local bias adjustment may help. PERSIANN offered moderate but stable performance, while GSMaP and AgERA5 require caution given their systematic underestimation and overestimation respectively. The authors also flag limitations: the 95-station network remains sparse relative to Ethiopia’s climatic heterogeneity, point-to-pixel comparisons introduce representativeness differences among products of different resolutions, land cover was mapped only for 2020 across the whole 24-year period, and extreme rainfall events were not directly verified, so flood-related conclusions await dedicated event-based testing. Future work, they suggest, should focus on elevation- and region-specific bias correction and on merging satellite estimates with gauge data to squeeze still more accuracy from Ethiopia’s most important and most elusive climate variable.

Subject of Research: Validation of six satellite precipitation products against rain gauge observations across Ethiopia

Article Title: Evaluation of satellite-based rainfall estimates and their performance across Ethiopia

Article References: Mulualem, G. M., Weldemariam, E. C., Orke, Y. A., Parsons, D., Bagiliko, J., Clements, L., & Stern, D. (2026). Evaluation of satellite-based rainfall estimates and their performance across Ethiopia. Theoretical and Applied Climatology, 157(11), Article 751. https://doi.org/10.1007/s00704-026-06679-w

Image Credits: AI Generated

DOI: 10.1007/s00704-026-06679-w

Keywords: satellite precipitation, Ethiopia, CHIRPS, IMERG, GSMaP, PERSIANN, AgERA5, rain gauge validation, rain-fed agriculture, drought monitoring, elevation, land cover