From data to decisions: ensemble deep learning framework for maize seed variety classification with explainable AI

Abstract

Accurate and non-destructive identification of maize seed varieties is essential for maintaining seed purity, improving crop productivity, and supporting agricultural quality assurance. Although recent studies report high accuracies using advanced imaging and deep learning techniques, many rely on hyperspectral sensors or controlled acquisition environments, or offer limited interpretability, thereby restricting real-world adoption. This study investigates maize seed variety classification using low-cost RGB imaging and an ensemble deep learning framework, focusing on three widely cultivated Ghanaian varieties: Bhihilifa, Sanzal Sima, and Wang Dataa. In collaboration with Heritage Seeds Ghana, an expert-verified dataset of 17,724 maize seed images was curated under realistic acquisition conditions with an Apple iPhone 12. A baseline convolutional neural network (CNN) was developed for benchmarking, followed by transfer learning with DenseNet169, MobileNetV2, and Xception architectures. An ensemble framework integrating DenseNet169 and Xception was proposed to improve robustness and generalization. Across five-fold cross-validation, the ensemble achieved a mean accuracy of 98.89±0.17%, precision of 98.86±0.18%, recall of 98.80±0.17%, and F1-score of 98.83±0.17%. To enhance transparency, Grad-CAM and saliency maps were employed to visualize discriminative features influencing model decisions. Results demonstrate that accurate and interpretable classification of maize seed varieties is achievable using standard RGB imagery without specialized sensing equipment.