Hirko et al. (2025) Using machine learning and satellite data to analyse climate change in the Upper Awash Sub-basin, Ethiopia

Identification

• Journal: Physics and Chemistry of the Earth Parts A/B/C
• Year: 2025
• Date: 2025-10-10
• Authors: Deme Betele Hirko, J A du Plessis, A. Bosman
• DOI: 10.1016/j.pce.2025.104137

Research Groups

• Stellenbosch University, Faculty of Engineering, Department of Civil Engineering, Stellenbosch, Western Cape, South Africa

Short Summary

This study integrates satellite and observational datasets with future projections using both machine learning (ML) and climate modeling to assess long-term climate trends in the Upper Awash Sub-basin, Ethiopia. It reveals significant disparities between ML-based projections, which indicate a slight precipitation rise and temperature decrease, and CMIP6 SSP5-8.5 projections, which anticipate a substantial precipitation decline and temperature increase, underscoring the need for multi-model and region-specific analyses.

Objective

• To evaluate the predictive performance of selected machine learning models in forecasting temperature and precipitation for the Upper Awash Sub-basin from 2025 to 2075.
• To compare these machine learning projections with outputs from CMIP6 models under the Shared Socioeconomic Pathway (SSP) scenario SSP5-8.5 (high emissions).

Study Configuration

• Spatial Scale: Upper Awash Sub-basin, Ethiopia (Akaki sub-basin, 1651 km², within the broader Awash River Basin, 114,123 km²).
• Temporal Scale:
  • Historical: 1948–2010 (Princeton, CMIP6), 1981–2010 (Ethiopian National Meteorology Agency).
  • Future Projections: 2025–2075 (Princeton ML, CMIP6 SSP5-8.5).

Methodology and Data

• Models used:
  • Machine Learning: Random Forest (RF), Extra Trees (ET), K-Nearest Neighbours (KNN), Long Short-Term Memory (LSTM), Light Gradient Boosting Machine (LGBM), Feed-forward Neural Networks (FNN). (RF was selected as the best-performing model for future projections).
  • Climate Models: Coupled Model Intercomparison Project Phase 6 (CMIP6) – specifically the Geophysical Fluid Dynamics Laboratory Climate Model version 4 (GFDL-CM4) under the SSP5-8.5 scenario.
  • Hydrological/Water Management: WEAP (Water Evaluation and Planning) modelling framework (version 2023.0).
• Data sources:
  • Satellite: Princeton Global Forcing Dataset (1948–2010, 28 km spatial resolution), European Space Agency Climate Change Initiative Land Cover (ESA-CCI-LC).
  • Observation: Ethiopian National Meteorology Agency (ENMA) historical monthly climate data (temperature and precipitation) for the Addis Ababa region (1981–2010) from Bole, Sebeta, Aleltu, Boneya, Sendefa, and Bulbula stations.
  • Reanalysis/Model Output: CMIP6 Climate Model – GFDL-CM4 (Historical: 1948–2010; Projection: 2025–2075), NASA Earth Exchange Global Climate Model (NEX-GCM) for downscaling and bias correction of CMIP6 data.
  • Base maps: USGS, OpenStreetMap, Google, National Geographic, and ESRI (integrated into WEAP).

Main Results

• The Random Forest (RF) machine learning model demonstrated the highest predictive accuracy for historical satellite data, achieving R² values of 0.97 for training and 0.96 for testing temperature, and 0.93 for training and 0.94 for testing precipitation.
• Historical comparisons (1948–2010 / 1981–2010) revealed that Princeton satellite data showed 6.3 % more precipitation and 10.2 % lower temperature than observed records. In contrast, CMIP6 historical data indicated 57.9 % less precipitation and an 18.2 % increase in temperature compared to observed records.
• For future projections (2025–2075), the Princeton ML model projects a slight 0.4 % rise in precipitation and a 10.2 % decrease in temperature (compared to observed RF ML model future projections).
• Conversely, the CMIP6 Shared Socioeconomic Pathway (SSP5-8.5) scenario anticipates a 52.1 % decline in precipitation and a 26.1 % rise in temperature (compared to observed RF ML model future projections).
• A direct comparison between the SSP5-8.5 scenario and Princeton ML projections shows a 52.3 % precipitation decrease and a 40.4 % temperature increase.
• These disparities highlight fundamental differences in model assumptions, spatial sensitivity, and data sources, with CMIP6 suggesting intensified warming and drying, while ML-based results indicate possible localized climate buffering effects.

Contributions

• This study is among the first to integrate satellite and observational datasets with future projections using both machine learning and climate modeling approaches for long-term climate trend assessment in the Upper Awash Sub-basin, Ethiopia.
• It provides a critical comparative analysis of ML and CMIP6 results, highlighting their respective strengths and discrepancies, which is crucial for supporting robust water resource planning in climate-vulnerable regions.
• The research underscores the importance of multi-model comparisons and region-specific analyses for developing effective climate adaptation strategies.
• It contributes to the growing field of hybrid climate modeling by demonstrating how artificial intelligence can enhance traditional approaches to climate risk analysis.
• The findings emphasize the necessity of incorporating longer-term and more comprehensive historical datasets for accurate machine learning climate predictions and advocate for improved inter-institutional cooperation in compiling and updating climate data.

Funding

• This research did not receive any funding.

Citation

@article{Hirko2025Using,
  author = {Hirko, Deme Betele and Plessis, J A du and Bosman, A.},
  title = {Using machine learning and satellite data to analyse climate change in the Upper Awash Sub-basin, Ethiopia},
  journal = {Physics and Chemistry of the Earth Parts A/B/C},
  year = {2025},
  doi = {10.1016/j.pce.2025.104137},
  url = {https://doi.org/10.1016/j.pce.2025.104137}
}