The project focused on the Kulfo River in southern Ethiopia, part of the Rift Valley where limited observations complicate efforts to assess climate-driven changes in water availability and floods. The study assessed 16 CMIP6 models under the SSP1-2.6, SSP2-4.5 and SSP5-8.5 scenarios. It then combined bias-corrected climate projections with a hydrological model and machine-learning prediction methods, comparing approaches that might be useful where long, high-quality records are scarce. After correction, CNRM-ESM2-1 produced the strongest precipitation and minimum-temperature results, while CNRM-CM6-1 performed best for maximum temperature, according to the reported model comparison.

Future projections showed progressively greater streamflow reductions by 2100, with annual discharge estimated to decline by approximately 8.5% under SSP1-2.6, 13.9% under SSP2-4.5 and 20.7% under SSP5-8.5. Temperatures were projected to increase in each scenario. The study associated the declining discharge with increased evaporative demand and altered seasonal rainfall, including reduced summer flow and shifts in hydrological seasonality. Conformal prediction was used to express uncertainty around the modeled outcomes, and the analysis found that this uncertainty increased in the projections, particularly under the high-emissions SSP5-8.5 scenario. A separate non-stationary flood-frequency analysis also indicated significant changes in future flood characteristics, although the supplied findings do not provide specific change magnitudes for individual flood metrics.

The results suggest that careful correction of climate-model biases can make global climate ensembles more useful for hydrological assessment in a data-limited basin. The reported validation score of 0.806 is lower than the calibration value of 0.999, illustrating that performance outside the training period was not identical to performance during fitting. The authors also found that Random Forest outperformed the tested Long Short-Term Memory and other deep-learning methods, which they interpreted as evidence that simpler ensemble approaches may be suitable when training data are limited. This does not establish that Random Forest will perform equally well in other catchments: the findings are model- and study-specific, and the study’s scope was the Kulfo watershed rather than Ethiopian rivers generally. Even so, the integrated framework links climate-model bias correction, streamflow prediction, uncertainty estimation and flood assessment, offering a structured way to examine future water-resource risks where direct historical observations are sparse.

Subject of Research: CMIP6 bias correction and machine-learning projections of streamflow change in Ethiopia’s Kulfo River watershed

Article Title: Machine learning and bias corrected CMIP6 for improved streamflow prediction in the data-scarce Kulfo river, Ethiopia

Article References: Ashagre, H. M., Goshime, D. W., Yersaw, B. T., & Walle, M. A. (2026). Machine learning and bias corrected CMIP6 for improved streamflow prediction in the data-scarce Kulfo river, Ethiopia. Discover Sustainability. https://doi.org/10.1007/s43621-026-04724-2

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04724-2

Keywords: Kulfo River, Ethiopia, CMIP6, streamflow projection, Random Forest, bias correction, climate change, hydrology, flood risk, prediction uncertainty, Machine, learning