mohamed et al. (2025) A hybrid deep learning and rule-based model for smart weather forecasting and crop recommendation using satellite imagery

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-10-15
• Authors: salma anwer mohamed, Olfat O. Abdel Maksoud, A. Fathy, Ahmed S.A. Mohamed, Khaled M. Hosny, Hatem M. Keshk, Sayed A. Mohamed
• DOI: 10.1038/s41598-025-21506-4

Research Groups

• Faculty of Computers Science and Technology, Modern Academy, Cairo, Egypt
• National Authority for Remote Sensing and Space Science, Cairo, Egypt

Short Summary

This study develops a hybrid deep learning and rule-based framework for smart weather forecasting and crop recommendation in Egypt's Al-Sharkia region, focusing on rice and wheat. The framework integrates CNN-based land suitability classification (training loss reduced from 0.2362 to 6.87e-4) with RNN-LSTM-based weather prediction (Root Mean Squared Error of 0.19) and rule-based crop advisories to provide precise, localized agricultural guidance.

Objective

• To develop an innovative hybrid framework integrating multispectral image analysis, advanced weather forecasting, and rule-based models to improve agricultural practices in Egypt's Al-Sharkia region, specifically for rice and wheat cultivation.
• To provide precise, localized forecasts and customized agricultural advice to facilitate informed decisions regarding crop selection, planting schedules, and resource allocation.
• To mitigate agricultural productivity declines, address land degradation, reduce crop losses and operational costs, and encourage sustainable agricultural practices in response to climate change.

Study Configuration

• Spatial Scale: Al-Sharkia Governorate, Northern Egypt (total area of 4911 km²).
• Temporal Scale:
  • CNN-based land suitability: 2022 to 2023 crop season.
  • RNN-LSTM-based weather forecasting: 2020 to 2024, with a prediction horizon of up to 1 year (short to medium-term, 1 to 5 days lead time for NOAA data).

Methodology and Data

• Models used:
  • Convolutional Neural Network (CNN) for spatial classification of agricultural land suitability.
  • Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) units for temporal weather forecasting.
  • Rule-based classifier for crop recommendation.
• Data sources:
  • Satellite imagery: Sentinel-2 Level-2A surface reflectance data, NASA Copernicus system API.
  • Remote sensing devices.
  • Meteorological stations (Egyptian).
  • National Oceanic and Atmospheric Administration (NOAA) outputs (including COSMO model) for weather forecasts.
  • Ground-truth data from field campaigns in Sharqiyah.
  • Derived spectral indices: Normalized Difference Vegetation Index (NDVI), Normalized Difference Water Index (NDWI), Moisture, Enhanced Vegetation Index (EVI), Normalized Difference Snow Index (NDSI), Green Normalized Difference Vegetation Index (GNDVI).
  • Meteorological variables: air temperature at 2 meters above ground level, surface skin temperature (°C), atmospheric pressure at station level (hPa), specific humidity (g/kg), relative humidity (%) at 2 meters above ground level, total precipitation (mm), dew point temperature at 2 meters above ground level, and mean wind speed at 10 meters above ground level.

Main Results

• CNN Model for Land Classification:
  • Achieved a substantial reduction in training loss from 0.2362 to 6.87e-4 by Epoch 100, with an assessment loss of 2.96.
  • Demonstrated high classification performance for agricultural land suitability: High Agricultural Preference (Precision 0.95, Recall 0.97, F1-Score 0.96), Low Agricultural Preference (Precision 0.92, Recall 0.89, F1-Score 0.90), Not Suitable for Agriculture (Precision 0.92, Recall 0.90, F1-Score 0.91).
• RNN-LSTM Model for Weather Forecasting:
  • Achieved a Root Mean Squared Error (RMSE) of 0.19, indicating high predictive precision.
  • Accurately represented temporal trends, exemplified by a temperature decline from 15.34 °C to 12.97 °C over 11 days (January 1st to January 11th).
  • Predicted a substantial reduction in relative humidity from 80.00% to 43.88% (January 1st to January 9th), followed by an increase, reflecting typical atmospheric phenomena.
  • Forecasted negligible precipitation and constant air pressure (approximately 100.87 hPa).
  • Showed significant quantitative improvement over baseline methods: 86.5% enhancement in forecasting accuracy compared to Singh et al. (2019) (RMSE 1.41) and almost 80% enhancement compared to Xu et al. (2024) (RMSE 0.95).
• Hybrid Framework (CNN + LSTM + Rule-based):
  • Ablation study confirmed the essential contribution of each component: removing CNN raised RMSE to 0.37, removing LSTM raised RMSE to 0.45, and removing the rule-based component resulted in a 22% decline in crop recommendation accuracy.
  • Provides customized advice on crop selection and timing, reducing crop failure likelihood and promoting sustainable agriculture.

Contributions

• Introduction of an innovative hybrid framework that amalgamates CNN-based image analysis, LSTM-based weather prediction, and rule-based crop advisories for precision agriculture.
• Development of a comprehensive method that provides precise, localized forecasts and customized agricultural advice, facilitating informed decisions regarding crop selection, planting schedules, and resource allocation.
• Strategic mitigation of agricultural productivity declines and land degradation by determining ideal planting periods for rice and wheat through dynamic analysis of temporal weather patterns and soil health using advanced deep learning techniques.
• Enhancement of agricultural decision-making by integrating real-time soil classifications and climatic forecasts with a database of ideal crop growth circumstances, leading to reduced annual farmer losses and optimized resource use.
• Validation of a scalable and cost-effective forecasting instrument for sustainable agriculture, with significant improvements in predictive accuracy over existing baseline methods.

Funding

• The Science, Technology & Innovation Funding Authority (STDF)
• The Egyptian Knowledge Bank (EKB)

Citation

@article{mohamed2025hybrid,
  author = {mohamed, salma anwer and Maksoud, Olfat O. Abdel and Fathy, A. and Mohamed, Ahmed S.A. and Hosny, Khaled M. and Keshk, Hatem M. and Mohamed, Sayed A.},
  title = {A hybrid deep learning and rule-based model for smart weather forecasting and crop recommendation using satellite imagery},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-21506-4},
  url = {https://doi.org/10.1038/s41598-025-21506-4}
}