2025, Vol. 10, Special Issue 12

Background: Accurate crop yield prediction is essential for food security, economic planning, and sustainable agriculture. Traditional statistical and machine learning methods often fail to capture the complex spatial-temporal interactions among climatic variables, soil characteristics, vegetation indices, and management practices.
Objective: This study proposes a Hybrid Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) deep learning framework to enhance crop yield prediction by integrating spatial and temporal agricultural datasets.
Methods: Multisource data-including MODIS and Sentinel-2 satellite imagery, NASA POWER and NOAA climate data, soil property maps, and ground-truth yield records were collected for five major crops over a 10-year period. Data pre-processing involved cleaning, normalization, temporal alignment, and image augmentation. CNN layers extracted spatial features from vegetation indices, while stacked LSTM layers modeled temporal dependencies from climatic and phenological sequences. Spatial and temporal embeddings were fused, followed by dense layers for yield estimation. The model was trained using the Adam optimizer with 100 epochs and validated through five-fold cross-validation. Performance was evaluated via RMSE, MAE, R², and prediction deviation.
Results: The hybrid CNN-LSTM model achieved superior performance (R²=0.92, RMSE=186.5 kg/ha, MAE=122.3 kg/ha), outperforming LSTM-only, CNN-only, Random Forest, and SVR models. NDVI, rainfall, and temperature emerged as the most influential features. Spatial heatmaps and learning curves confirmed high model accuracy and robust convergence. Statistical tests indicated no significant difference between predicted and actual yields (p>0.05).
Conclusion: The proposed CNN-LSTM framework effectively integrates spatial and temporal agricultural data to deliver reliable, high-precision crop yield predictions. Its practical applicability supports precision agriculture, resource planning, and climate-smart farming strategies.