📄 earth_science

Machine Learning and Bias-Corrected CMIP6 for Improved Streamflow Prediction in the Data-Scarce Kulfo River, Ethiopia

This study demonstrates that integrating bias-corrected CMIP6 climate projections with machine learning models, particularly Random Forest, significantly improves streamflow prediction accuracy in the data-scarce Kulfo River basin, revealing a future trend of rising temperatures and declining annual streamflow with increased hydrological uncertainty under various climate change scenarios.

Original authors: Hailemariam Molla Ashagre, Demelash Wondimagegnehu Goshime, Babur Tesfaye Yersaw, Melaku Adugnaw Walle

Published 2026-06-29

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Hailemariam Molla Ashagre, Demelash Wondimagegnehu Goshime, Babur Tesfaye Yersaw, Melaku Adugnaw Walle

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Predicting the River's Future in a Data-Desert

Imagine the Kulfo River in Ethiopia as a vital lifeline for a bustling town and its farms. It's like the main artery of a body, carrying water to keep everything alive. However, scientists face a problem: they don't have enough historical records (data) to know exactly how this river behaves, and the weather is changing fast.

This study is like a team of detectives trying to predict the river's future behavior using two powerful tools:

Super-Computer Weather Models (CMIP6): These are like giant, global weather simulators that try to guess what the climate will look like in 2100.
Machine Learning (AI): This is like a smart student who learns from past patterns to make better guesses than a human could.

The goal? To figure out if the Kulfo River will have more water, less water, or more floods in the future, so the local community can prepare.

Step 1: Fixing the "Broken" Weather Simulators

The researchers started by looking at 16 different global weather models. Think of these models as 16 different weather forecasters.

The Problem: Before they were "fixed," these forecasters were terrible at predicting the local weather. Some said it would rain buckets when it was dry; others said it would be freezing when it was hot. They were like a group of friends guessing the temperature, but everyone was off by a huge margin.
The Fix: The team used a "bias correction" technique. Imagine taking those 16 friends and giving them a calibration tool (a specific set of rules based on real local observations) to adjust their guesses.
The Result: After the fix, the forecasters became incredibly accurate. One model, CNRM-ESM2-1, became the "star student" for predicting rain and cold nights, while another, CNRM-CM6-1, became the best at predicting hot days. The error in their rain predictions dropped by about 92%—like going from guessing the wrong season entirely to guessing the exact hour of the day.

Step 2: Teaching the AI to Read the River

Next, the team needed to figure out how the river reacts to this weather. They tried teaching several different types of "AI students" (Machine Learning models) to predict the river's flow based on rain and temperature.

The Contest: They pitted different algorithms against each other, including complex deep learning models (like LSTM) and simpler, robust models.
The Winner: The Random Forest model won the competition. Think of Random Forest as a committee of 500 experts who vote on the answer. Instead of relying on one opinion, they average out their guesses.
Why it won: This "committee" was incredibly accurate. During its training, it got a near-perfect score (99.9% accuracy). Even when tested on new data it hadn't seen before, it still performed very well (80.6% accuracy). It learned that the river's flow depends heavily on what happened the day before (like a river that remembers the rain from yesterday).

Step 3: The Future Forecast (What Happens to the River?)

Using the "fixed" weather models and the "winning" AI, the team looked at three different future scenarios (from a "green" future to a "high pollution" future). Here is what they found:

The Heat is On: No matter which future scenario happens, the area is getting hotter. Both the hottest days and the coldest nights are warming up. By the end of the century, under the worst pollution scenario, the heat will be significantly higher than today.
The Rain is Staying Put (Mostly): The total amount of rain isn't changing drastically, but the timing is shifting.
The River is Shrinking: This is the big worry. Even though rain might stay similar, the river flow is predicted to drop significantly.
- In a "best case" scenario, the river might shrink by about 8.5%.
- In a "worst case" scenario, the river could shrink by nearly 21%.
- The Analogy: Imagine a bathtub where the faucet (rain) is running at the same speed, but the drain (heat/evaporation) is opening wider. The water level in the tub (the river) drops because more water is evaporating into the air before it can flow downstream.

Step 4: The Seasonal Shuffle

The study found that the river is changing its "personality" throughout the year:

Summer (The Wet Season): The river is getting much drier during the main rainy months. This is bad for farmers who rely on summer floods to water their crops.
Winter (The Dry Season): Surprisingly, the river might get a little bit more water in winter, but this doesn't make up for the massive loss in summer.
The Result: The river is becoming less reliable. It's like a friend who used to show up on time for a big party (summer) but is now skipping it, only showing up briefly for a coffee (winter).

Step 5: Uncertainty and Floods

The "Fog of War": The researchers used a special math trick called "Conformal Prediction" to draw a safety net around their guesses. They found that as we get further into the future (towards 2100), the "fog" gets thicker. The predictions become less certain, especially in the worst pollution scenarios.
The Flood Paradox: Here is a tricky part. Even though the average amount of water in the river is going down, the floods are getting more dangerous.
- The Analogy: Imagine a highway. Usually, there is steady traffic (steady river flow). Climate change is causing the traffic to stop and start erratically. Even if the total number of cars on the road decreases, the cars that do show up are driving much faster and crashing more often.
- The study found that while the river gets smaller overall, the "100-year flood" (a massive flood event) is becoming more frequent and intense. A flood that used to happen once every 100 years might happen much more often, or the water level during a flood could be lower but the speed and suddenness of the water could be more dangerous.

The Bottom Line

This paper tells us that the Kulfo River is facing a tough future.

The tools work: We can now use AI and corrected weather models to make very accurate predictions even in places where we don't have many data records.
The river is shrinking: Due to rising heat, the river will likely have less water overall, especially in the summer.
The danger is shifting: While the river gets smaller, the risk of sudden, intense floods is changing, and the timing of the water is becoming unreliable for farming and daily life.

The authors suggest that to survive this, the community needs to plan for a river that is smaller, hotter, and more unpredictable, using these new AI tools to help make smart decisions about water management.

Technical Summary: Machine Learning and Bias-Corrected CMIP6 for Improved Streamflow Prediction in the Data-Scarce Kulfo River, Ethiopia

1. Problem Statement

Climate change is increasingly altering hydrological processes and water availability, posing critical challenges for data-scarce regions such as Ethiopia's Rift Valley basins. The Kulfo watershed in southern Ethiopia, which supports agricultural livelihoods, the Nech Sar National Park, and downstream communities along Lake Abaya, lacks comprehensive hydrological assessments that integrate state-of-the-art climate projections with advanced modeling techniques.

Traditional process-based hydrological models (e.g., SWAT) often struggle in developing country contexts due to the requirement for extensive parameterization and calibration data that are frequently unavailable. Furthermore, Global Climate Models (GCMs) from the Coupled Model Intercomparison Project Phase 6 (CMIP6) exhibit systematic biases when applied at local scales, necessitating rigorous bias correction and downscaling before hydrological application. While machine learning (ML) offers promising alternatives for capturing non-linear climate-streamflow relationships, limited studies have systematically evaluated multiple ML algorithms for streamflow projection in Ethiopian basins using the latest CMIP6 scenarios.

2. Methodology

Study Area and Data

The study focuses on the Kulfo watershed (5.9°–6.2° N, 37.5°–37.8° E), a perennial river draining from the Gughe Highlands into Lake Abaya. The research utilized:

Observed Data: Daily precipitation and temperature records (1986–2014) from the Arba Minch meteorological station, and daily streamflow records (1980–2022) from the Ministry of Water and Energy gauging station.
Climate Data: Sixteen (16) CMIP6 global climate models were selected based on variable availability and spatial resolution. Projections were generated under three Shared Socioeconomic Pathways (SSP1-2.6, SSP2-4.5, and SSP5-8.5) for the period 2015–2100.

Bias Correction

To address systematic biases in raw CMIP6 outputs, the study applied both statistical and machine learning-based bias correction techniques. The models were calibrated using the 1986–2006 period and validated against 2007–2014 observations. Performance was evaluated using Root Mean Square Error (RMSE), Percent Bias (PBIAS), Correlation Coefficient (R), and Kling–Gupta Efficiency (KGE).

Machine Learning Model Development

The rainfall–runoff relationship was conceptualized as a function of current and antecedent hydro-climatic variables ( $P_t, P_{t-1}, P_{t-2}, T_{max,t}, T_{min,t}, Q_{t-1}$ ). Eight machine learning algorithms were developed and compared:

Ensemble Methods: Random Forest (RF), Bagging, Boosting (GBM), AdaBoost, CatBoost, and Stacking.
Other Algorithms: Long Short-Term Memory (LSTM), K-Nearest Neighbor (KNN), and Support Vector Machine (SVM).

Hyperparameters were optimized for each model. The dataset was partitioned into a training period (1986–2006, 70%) and a testing/validation period (2007–2014, 30%). Model performance was assessed using Nash–Sutcliffe Efficiency (NSE), Coefficient of Determination ( $R^2$ ), RMSE, and Mean Absolute Error (MAE).

Projection and Uncertainty Framework

The best-performing ML model was used to simulate daily streamflow using bias-corrected CMIP6 inputs. The study incorporated two advanced analytical frameworks:

Conformal Prediction (CP): Integrated with the Random Forest model to quantify uncertainty in future discharge projections, generating 95% prediction intervals (PI) to assess the reliability and spread of estimates.
Non-Stationary Flood Frequency Analysis: A time-varying Generalized Extreme Value (GEV) model was fitted to annual maximum discharge series, where the location parameter was expressed as a function of projected temperature anomalies ( $\mu(t) = \mu_0 + \mu_1 \cdot \Delta T(t)$ ). This approach allowed flood probabilities and return periods to evolve over time rather than remaining constant.

3. Key Results

CMIP6 Model Performance and Bias Correction

Raw CMIP6 models exhibited substantial systematic biases, with precipitation RMSE values up to 2.95 mm/day and negative KGE values for many models.

Bias Correction Efficacy: The application of bias correction (specifically Random Forest for precipitation) significantly improved model reliability. Precipitation RMSE was reduced by an average of 83% (up to 92% for specific models like MIROC6), with post-correction correlation coefficients exceeding 0.97.
Best Performing Models:
- Precipitation: CNRM-ESM2-1 and CNRM-CM6-1 showed the best agreement with observations after correction.
- Maximum Temperature: CNRM-CM6-1 emerged as the top performer (RMSE = 0.24°C, KGE = 0.69).
- Minimum Temperature: CNRM-ESM2-1 demonstrated the highest accuracy (RMSE = 0.09°C, KGE = 0.72).

Machine Learning Model Comparison

Among the eight algorithms tested, Random Forest (RF) demonstrated superior predictive capability:

Calibration: NSE = 0.999, $R^2$ = 0.9995.
Validation: NSE = 0.806.
Variable Importance: Antecedent streamflow ( $Q_{t-1}$ ) was identified as the dominant predictor (42% importance), reflecting significant groundwater contributions to the watershed.
Other ensemble methods (AdaBoost, Boosting) performed well, while LSTM, KNN, and SVM showed weaker predictive capabilities, likely due to the limited training period and data sparsity.

Future Climate and Streamflow Projections

Temperature: All SSP scenarios project consistent warming. By the late century (2071–2100), maximum temperatures are projected to exceed 32°C and minimum temperatures to rise above 18.5°C under high-emission scenarios (SSP5-8.5).
Precipitation: Precipitation is projected to remain relatively stable with modest declines (up to -5.1% under SSP5-8.5 by late century), though seasonal redistribution is significant.
Annual Streamflow: Future projections indicate a progressive decline in annual streamflow across all scenarios. By the end of the century, reductions are projected to reach:
- -8.5% under SSP1-2.6.
- -13.9% under SSP2-4.5.
- -20.7% under SSP5-8.5.
Seasonal Redistribution: A significant shift in hydrological seasonality is observed, characterized by substantial reductions in summer discharge (up to -33.3% under SSP5-8.5) and moderate increases in winter flow (initially +21%, though diminishing over time).

Uncertainty and Flood Frequency

Conformal Prediction: Uncertainty in streamflow projections increases toward the end of the century, particularly under SSP5-8.5, with wider 95% prediction intervals indicating greater variability in hydrological responses.
Non-Stationary Floods: Contrary to the decline in mean flow, the non-stationary GEV analysis revealed a complex response in flood characteristics. The 100-year flood return level ( $Q_{100}$ ) is projected to decrease significantly (by 35–41% by 2100) due to reduced baseflow from higher evapotranspiration. However, the study notes a counterintuitive finding where extreme flood magnitudes may intensify due to more intense, short-duration rainfall events, shortening the 50-year flood return period to 8 years by 2080.

4. Significance and Contributions

This study provides the first comprehensive assessment of CMIP6 model performance in the Kulfo watershed and demonstrates the superiority of ensemble machine learning approaches, specifically Random Forest, for streamflow projection in data-scarce Ethiopian catchments.

The research makes several specific contributions:

Methodological Integration: It successfully integrates bias-corrected CMIP6 projections with machine learning to address data limitations, offering a robust framework for future streamflow estimation.
Uncertainty Quantification: By employing conformal prediction, the study quantifies the uncertainty in future discharge estimates, highlighting the increasing variability under high-emission scenarios.
Non-Stationary Analysis: The application of a time-varying GEV model challenges stationary assumptions, revealing that climate change may alter flood characteristics in complex ways (reducing mean flow while potentially intensifying flash floods).
Policy Relevance: The findings underscore the vulnerability of the Kulfo watershed to future hydro-climatic changes. The study emphasizes the need to integrate machine learning, uncertainty analysis, and non-stationary hydrological approaches into climate-resilient water resource planning to support sustainable watershed management, irrigation development, and flood risk reduction in Ethiopia and similar tropical catchments.

The authors conclude that while the study offers valuable insights, future research should focus on expanding hydro-meteorological monitoring networks, integrating physically based models with hybrid deep learning approaches, and incorporating socioeconomic drivers to further enhance predictive robustness.