Comparative Evaluation of AI, Physics, and Hybrid Models for Daily Rainfall Prediction in Semi-Arid Katsina, Nigeria
This study compares artificial intelligence, physics-based, and hybrid modeling approaches for daily rainfall prediction in semi-arid Katsina. It shows how physical interpretability and machine-learning flexibility can be combined to improve forecasting performance in data-limited climate contexts.
Abstract Summary
Accurate rainfall prediction remains difficult in semi-arid regions because of high temporal variability, limited observation networks, and localized convective rainfall patterns. The study benchmarks an AI model, a physics-based model, and a hybrid model using daily rainfall data for Katsina, Nigeria.
The hybrid model achieved the best overall performance by combining physically derived predictors with machine-learning capability, making it more accurate and more useful for operational forecasting in resource-constrained settings.
Model Performance
The study compares model skill across statistical and hydrological metrics, with the hybrid approach outperforming both standalone AI and physics-based approaches.
Methods
- AI modeling used tree-based machine-learning methods including Random Forest and XGBoost.
- Physics-based modeling used physically derived atmospheric proxies and non-negative least squares.
- Hybrid modeling tested residual learning, feature augmentation, and blending strategies.
- Evaluation used training data from 2014-2020, validation data from 2021-2022, and testing data from 2023-2024.
Why It Matters
Improved rainfall forecasting supports drought and flood early warning, agricultural planning, water management, and climate-resilience decisions in northern Nigeria. The study provides a practical model for advancing locally relevant climate intelligence systems across data-scarce African regions.