Sonon et al. (2026) Artificial intelligence methods for rainy season forecasting: a comprehensive analysis

Identification

• Journal: Modeling Earth Systems and Environment
• Year: 2026
• Date: 2026-03-30
• Authors: Bienvenu Sonon, Charlemagne D. S. J. Gbemavo, Romain Glèlè Kakaï
• DOI: 10.1007/s40808-026-02769-w

Research Groups

• Laboratoire de la Biomathématique et d’Estimations Forestières, Université d’Abomey-Calavi, Cotonou, Benin

Short Summary

This critical review comprehensively analyzes artificial intelligence methods for rainfall forecasting, focusing on data validation, prediction across various time horizons, and key rainy season parameters, revealing hybrid models as the most prevalent approach. The study proposes a novel classification framework for AI models based on forecast horizon, target parameters, and algorithmic complexity, while highlighting gaps in methodological standardization and advanced AI model adaptation in tropical regions.

Objective

• To synthesize trends in artificial intelligence (AI) forecasting techniques applied to rainfall forecasting (both amount and occurrence) worldwide.
• To evaluate the reliability of these rainfall predictions across different climatic and geographic settings.
• To address the knowledge gap regarding the most widely used AI methods for forecasting rainy season characteristics (onset, cessation, duration) in diverse global regions.
• To analyze and discuss the accuracy of rainfall forecasts and identify differences between forecasting techniques, considering data validation, prediction methods over different time horizons, and specific rainfall parameters.

Study Configuration

• Spatial Scale: Global (worldwide), with a significant focus on Asia (55.89% of studies, particularly China, India, Iran, Malaysia), Australia (7.65%), South America, and North America. Africa and Europe are comparatively underrepresented. Specific mentions of tropical and sub-Saharan contexts.
• Temporal Scale: The review systematically examined AI applications between 2010 and 2024. Forecast horizons include hourly, daily, weekly, monthly, seasonal, and annual scales, as well as specific parameters like the onset, cessation, and duration of the rainy season.

Methodology and Data

• Models used:
  • Dominant: Hybrid models (53.22%, including NWP-AI, AI-AI, Statistics-AI combinations), Long Short-Term Memory (LSTM, 9.14%), Convolutional Neural Networks (CNN, 4.84%), Random Forest (RF, 4.3%).
  • Other frequently used: Artificial Neural Networks (ANN), Recurrent Neural Networks (RNN), Support Vector Machines (SVM), Wavelet Artificial Neural Network (WANN), Time-Delay Neural Network (TDNN), Echo State Networks (ESN), Extreme Learning Machine (ELM).
  • Best performing (≥ 90% precision): Adaline-AI, Bidirectional LSTM (BLSTM), Bayesian Neural Networks (BNN), CatBoost, ConvLSTM, DEEPRANCE-U-Net, Non-linear-AutoCNN, Decision Trees (DT), SVM, RNN, Ensemble Multiscale Gaussian Processes (EMSGP).
  • Emerging (rarely applied to seasonal parameters): PCA-UNet (Attention U-Net), Diffusion-based generative models (e.g., GED), Encoder-Decoder CNN, Graph Neural Networks (GNNs, e.g., HDGN), Generative Adversarial Networks (GANs, e.g., TU2Net-GAN, TS-RainGAN), Autoregressive Transformers (e.g., ST-TransNet).
• Data sources:
  • Observational data (48.90% of studies), often from meteorological stations or in situ measurements.
  • Hybrid approach combining observational and satellite data (12.64%).
  • Satellite data (3.83%).
  • Exogenous variables (e.g., large-scale climate indicators like ENSO, ocean oscillations) were incorporated in 44.21% of studies, while 37.91% focused solely on endogenous variables (e.g., historical rainfall data).

Main Results

• The review analyzed 183 publications from 2010 to 2024, revealing a growing interest in AI for rainfall forecasting, peaking in 2023 with 40 articles.
• Hybrid models are the most commonly used (53.22%), particularly NWP-AI combinations (47.36%), followed by LSTM (9.14%), CNN (4.84%), and RF (4.3%).
• The primary objectives for using AI in rainfall forecasting are water resource management (43%), water resource management for agriculture (19.8%), flood control and prevention (16.3%), and disaster mitigation and prevention (14.5%).
• Over half of the studies (55.89%) were conducted in Asia, with China (22.35%) and India (21.18%) being major contributors.
• Quantitative rainfall forecasting is dominant, with monthly (31.32%) and seasonal (18.13%) horizons being most frequent. Binary forecasting (rain/no rain) is also significant, especially for hourly (10.99%) and daily (4.95%) predictions.
• Forecasting of critical rainy season parameters (onset, cessation, duration) is underrepresented, accounting for only 1.65% of studies, primarily using RNN, Echo State Networks, and TDNN.
• Common data validation techniques include Exploratory Data Analysis (43.13%), Decomposition methods (33.33%, e.g., wavelet transformation), and Advanced data Validation Techniques (23.53%, e.g., cross-validation).
• Best-performing models (≥ 90% precision) include advanced AI models (e.g., Adaline-AI, BLSTM, ConvLSTM) and widely used models (e.g., DT, SVM, RNN), indicating that frequent usage does not always equate to superior performance.
• Key evaluation metrics for quantitative forecasts are RMSE (95 occurrences), R (52 occurrences), and MAE (46 occurrences). For binary forecasts, CSI (11 occurrences), FAR (10 occurrences), and POD (9 occurrences) are prevalent.
• Emerging AI paradigms (e.g., GANs, GNNs, Transformers) are rarely applied to seasonal rainfall parameters like onset, cessation, and duration, despite their success in short-term or binary predictions.
• Common hyperparameters include RBF (29.23%), ReLU (24.62%), and Sigmoid (20%) as activation functions; ADAM (36.67%) and Gradient Boosting (23.33%) as optimization algorithms; and Grid Search (47.06%) for hyperparameter tuning.

Contributions

• Provides a comprehensive, integrative, and quantitative synthesis of AI applications in rainfall forecasting, extending beyond conceptual or geographically restricted previous reviews.
• Systematically evaluates AI model usage across multiple temporal horizons (daily, weekly, monthly, seasonal, annual) and diverse climatic regions, including tropical and sub-Saharan contexts.
• Proposes a novel classification framework for AI models based on forecast horizon, target rainfall parameters, and algorithmic complexity.
• Offers new benchmarking insights into the distribution and performance of AI techniques across different timescales and climatic zones.
• Highlights the critical underrepresentation and methodological inconsistencies in forecasting agriculturally significant rainy season parameters (onset, cessation, duration).
• Emphasizes the need for standardization of methodologies for seasonal rainfall parameters and greater adaptation of advanced AI models in tropical contexts.
• Outlines a forward-looking research agenda advocating for the explicit integration of critical rainy season parameters and the incorporation of Explainable AI (XAI) techniques (e.g., SHAP, LIME) to enhance model interpretability and trust.
• Aligns findings with United Nations Sustainable Development Goals (SDGs 2, 13, 15) to promote climate-resilient and sustainable agricultural practices.

Funding

• The study was supported by a DAAD scholarship.

Citation

@article{Sonon2026Artificial,
  author = {Sonon, Bienvenu and Gbemavo, Charlemagne D. S. J. and Kakaï, Romain Glèlè},
  title = {Artificial intelligence methods for rainy season forecasting: a comprehensive analysis},
  journal = {Modeling Earth Systems and Environment},
  year = {2026},
  doi = {10.1007/s40808-026-02769-w},
  url = {https://doi.org/10.1007/s40808-026-02769-w}
}