Hybridisation of feature engineering and deep learning for accurate and efficient classification of maize crop diseases and pests

Abstract In Sub-Saharan Africa, maize is a dietary staple. However, maize crop pests and diseases threaten harvests and undermine food security. Machine learning (ML) and deep learning (DL) have been widely used for the classification of these diseases. With sufficient data, DL methods generally outperform ML approaches. However, feature engineering can yield better results with limited data for models trained from scratch. Pre-trained DL models and data augmentation address data scarcity, yet these models often lack flexibility and can be resource-intensive. This paper proposes a dual-axis integration of feature learning and feature engineering for the classification of maize crop diseases and pests under conditions of severe class-level data scarcity. Validation was conducted using real-world datasets from Burkina Faso, Namibia, Uganda, and Tanzania. Two model variants were developed: an unpruned model incorporating a complete set of engineered features and a pruned model retaining only the most significant engineered features. Their performance was compared against 28 pre-trained models commonly used for crop disease classification in the literature, using accuracy, F1-score, parameter efficiency, and computational efficiency. Normalised, unweighted average scores showed the three highest-ranking models as the proposed unpruned model (0.9742), MobileNetV3Small (0.9736), and MobileNetV3Large (0.9731). EfficientNetV2L had the lowest score of 0.4772. Cross-regional evaluation of the proposed models using an independent maize-disease dataset from Bangladesh indicated that their convolutional representations, learned from the Sub-Saharan African dataset, retained useful discriminative information after retraining only the classification head. Deployment-relevant metrics revealed that MobileNetV3Small offers the fastest inference (2.01 ms) and lowest latency-derived energy estimate (30.09 mJ), while the pruned model offers the best storage efficiency (2.24 MB). Thus, in resource-constrained settings, hybridising feature engineering and feature learning can yield more effective, deployable crop disease classification models.