Senjaliya et al. (2026) A comprehensive review of machine learning and deep learning approaches for rainfall forecasting: current progress, challenges, and future directions

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2026
• Date: 2026-02-28
• Authors: Jignesh Senjaliya, Vibhisha Vaghasia, Sanjay M. Shah
• DOI: 10.1007/s00704-026-06051-y

Research Groups

• Dept. of Computer Science, Kadi Sarva Vishwavidyalaya, Gandhinagar, Gujarat, India
• AnanthaQ Pvt. Ltd, Surat, Gujarat, India
• S.K. Patel Institute of Management & Computer Studies, Kadi Sarva Vishwavidyalaya, Gandhinagar, Gujarat, India

Short Summary

This comprehensive review synthesizes nearly 150 studies to compare statistical, machine learning (ML), and deep learning (DL) approaches for rainfall forecasting, identifying current progress, persistent challenges, and outlining future research directions for robust and climate-aware prediction systems.

Objective

• To evaluate the progress of ML and DL models for rainfall forecasting across different temporal and spatial scales.
• To examine hybrid frameworks and intelligent data fusion approaches that enhance predictive accuracy.
• To compare the relative strengths and limitations of statistical, ML, and DL frameworks.
• To identify ongoing challenges, such as data scarcity, non-stationarity, and model interpretability.
• To outline emerging opportunities, including transformer-based architectures, explainable AI, and integration with climate change projections.

Study Configuration

• Spatial Scale: Covers various spatial scales, from localized events to regional and global patterns, as a comprehensive review of existing literature.
• Temporal Scale: Encompasses diverse temporal scales, including hourly, daily, monthly, and seasonal rainfall predictions, as a comprehensive review of existing literature.

Methodology and Data

• Models used: Statistical models (ARIMA, GARCH, TAR), Machine Learning (Support Vector Machines, Random Forests, Gradient Boosting, k-Nearest Neighbors, Artificial Neural Networks, Ensemble/Hybrid ML), Deep Learning (Recurrent Neural Networks, Long Short-Term Memory networks, Gated Recurrent Units, Convolutional Neural Networks, Hybrid DL), and Advanced Architectures (Attention mechanisms, Transformers).
• Data sources: Satellite imagery, radar measurements, ground-based meteorological station observations, gridded climate datasets, and climate model outputs.

Main Results

• Traditional statistical models are foundational but are constrained by assumptions of linearity and stationarity, limiting their ability to capture non-linear dynamics and extreme events.
• Machine learning methods improved non-linear modeling but often require extensive feature engineering, are computationally expensive for large datasets, and exhibit inconsistent performance across different climatic regions and seasons.
• Deep learning approaches (RNNs, LSTMs, GRUs, CNNs) have significantly advanced rainfall prediction through automatic feature extraction and spatiotemporal learning, generally outperforming statistical and conventional ML methods. However, they are data-hungry, computationally demanding, prone to overfitting, and suffer from a lack of interpretability.
• Attention mechanisms and transformer architectures represent a new frontier, demonstrating superior ability to capture long-range dependencies and offering parallelized training. Preliminary studies show them outperforming conventional DL models, particularly for monthly rainfall prediction, though their application in hydrology is still nascent.
• Hybrid frameworks and intelligent data fusion techniques, combining heterogeneous data sources with ML/DL models, enhance robustness and adaptability, especially under changing climate conditions.

Contributions

• Provides a comprehensive synthesis of nearly 150 research studies on ML and DL approaches for rainfall forecasting, evaluating their progress across various scales.
• Systematically compares the strengths and limitations of statistical, ML, and DL frameworks, including emerging transformer-based models.
• Identifies critical ongoing challenges in rainfall forecasting, such as data scarcity, non-stationarity, and model interpretability.
• Outlines key future research directions, including hybrid frameworks, advanced architectures (transformers), explainable AI (XAI), data fusion, transfer learning, and operational real-time systems, offering a roadmap for future development.

Funding

This work received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.

Citation

@article{Senjaliya2026comprehensive,
  author = {Senjaliya, Jignesh and Vaghasia, Vibhisha and Shah, Sanjay M.},
  title = {A comprehensive review of machine learning and deep learning approaches for rainfall forecasting: current progress, challenges, and future directions},
  journal = {Theoretical and Applied Climatology},
  year = {2026},
  doi = {10.1007/s00704-026-06051-y},
  url = {https://doi.org/10.1007/s00704-026-06051-y}
}