Development of an Improved Hybrid Deep Learning Model for Cassava Yield Prediction

Abstract

Artificial Intelligence (AI) has significantly transformed the agricultural sector by enhancing crop yield forecasting, decision-making, and the analysis of soil and environmental conditions. Consequently, many studies have adopted deep learning (DL) techniques such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks for crop yield prediction. CNNs are effective in extracting spatial features from input data, while LSTMs capture temporal and phenological patterns, both of which are critical for accurate yield estimation. This study proposes a novel hybrid framework for cassava yield prediction that integrates CNN, LSTM, and Harris Hawk Optimisation (HHO). Soil and meteorological datasets were obtained from SoilGrids250m version 2.0 and the Nigerian Meteorological Agency (NIMET), Abuja, respectively, while cassava yield data were sourced from the Food and Agriculture Organisation Statistics. The datasets were preprocessed through noise removal, normalisation, and validation using a 5-fold cross-validation technique. The proposed HHO-CNN-LSTM model was evaluated against a conventional CNN-LSTM model using R-square (R²), mean absolute error (MAE), mean squared error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE). Results showed that the CNN-LSTM model performed poorly, with R² = −0.56726, MAE = 1.431207, MSE = 2.050261, RMSE = 1.431075, and MAPE = 13.54528. In contrast, the HHO-CNN-LSTM model achieved superior performance, recording R² = 0.370277, MAE = 0.904463, MSE = 0.82379, RMSE = 0.907629, and MAPE = 8. 745989.These findings demonstrate that integrating HHO significantly improves prediction accuracy and model reliability, making the proposed approach a valuable decision-support tool for sustainable cassava production and agricultural resilience.