Deep Learning Framework for Maize Demand Prediction Using Optimized Environmental Features

Abstract

Accurate maize demand prediction plays an important role in ensuring food security, market stability, and sustainable agricultural development. However, maize demand prediction is a complex process is mainly determined by the interactions between different factors. Therefore, traditional statistical models are ineffective due to their inability to capture the underlying nonlinear relationships and temporal dependencies in the data. It leads to resource mismanagement, overproduction, and underproduction that affects the economy. This work presents a novel approach based on Convolutional Neural Network-Bidirectional Long Short-Term Memory with Attention and Grey Wolf Optimized Environmental Features for maize demand prediction. Agricultural data such as soil, climate, irrigation, and production data were collected and pre-processed using the Multiple Imputation, Median Filtering, and Min-Max normalization to fill and smooth missing data. Then Grey Wolf Optimization is used for feature selection, which determines the relevant environmental features by removing extra ones and decreasing the complexity of the input data. The acquired features are processed through the Convolutional Neural Network (CNN) that identifies feature detectors, and their representations are extracted through BiLSTM with Attention, which means processing data in both directions. Moreover, the Attention mechanism highlights the important features that affect maize demand. After that, the production demand is predicted using the attained results for efficient planning and allocation of resources. The proposed framework achieved an RMSE of 2.14, MAE of 1.48, MAPE of 6.12%, and R² of 98.96%, that shows system's effectiveness in captures relevant nonlinear relationships and influential agricultural factors, resulting in more accurate maize demand prediction.