Prediction of Maize Yield in Uganda using CNN-LSTM Architecture on a Multimodal Climate and Remote Sensing Dataset

Accurate forecasting of maize yields is crucial for enhancing agricultural productivity and ensuring food security in Uganda. Traditional statistical methods for estimating crop yields face challenges in accuracy and scalability due to poor integration of diverse inputs and their inability to model the complex, nonlinear, and spatiotemporal aspects of crop growth dynamics. Consequently, this study developed a convolutional neural network and long short-term memory (CNN-LSTM) model to predict maize yields by leveraging biannual remotely sensed data and maize yield labels from the Zonal Agricultural Research and Development Institute (ZARDI) zones in Uganda. The dataset, covering the period from 2018 to 2020, includes satellite observations of climatic variables and vegetation indices. Although acquiring large amounts of satellite data for maize yield prediction is easy, accessing high-quality yield records across ZARDI zones remains challenging due to high costs and the time required. Thus, synthetic data augmentation using the Synthetic Minority Oversampling Technique for Regression (SMOGN) and feature dimensionality reduction based on the importance analysis of features was employed to increase and balance the yield dataset. The CNN-LSTM model’s ability to select features and perform extensive hyperparameter tuning enabled it to outperform baseline models. It achieved a Mean Squared Error (MSE) of 0.107 tonnes², a Mean Absolute Error (MAE) of 0.267 tonnes, a Root Mean Squared Error (RMSE) of 0.327 tonnes, and an R² score of 0.78. A comparative analysis showed that the CNN + Random Forest (RF) achieved an MSE of 0.137 tonnes ² , a MAE of 0.281 tonnes, an RMSE of 0.370 tonnes, and an R ² score of 0.722. These results outperformed the standalone CNN (MSE = 0.216, R ² = 0.562) and RF (MSE = 0.211, R ² = 0.573) models, underscoring the advantage of combining spatial-temporal learning for improved predictive accuracy. Residual analysis demonstrated the robustness of the proposed model, featuring minimal bias and an excellent fit between the actual and predicted yield. These findings highlight the potential of integrating deep learning and traditional machine learning for crop yield forecasting in diverse smallholder farming systems. Future research will focus on further integrating CNNs with Transformer architectures to enhance predictive accuracy and robustness. This provides a valuable framework for data-driven agricultural planning and decision-making in Uganda's ZARDI zones.