A Compact Deep Learning Architecture for Multi-Label Classification of Plant Diseases

Early and accurate detection of plant diseases, including cases of multiple co-occurring issues on the same plant, is essential for crop management and food security. This paper proposes a Compact Deep Learning Architecture (CDLA) optimized for multi-label classification of tomato leaf diseases and deficiencies. The model is built using depthwise separable convolutions and residual connections to achieve high accuracy with low computational cost, making it suitable for edge devices and real-time field deployment. We curated a balanced image dataset from Tomato Village and PlantVillage sources, encompassing 12 classes (single diseases and combinations of diseases / pest damage / nutrient deficiencies) to reflect real-world multi-pathogen scenarios. The CDLA model was trained with extensive data augmentation to enhance generalization. Evaluation results demonstrate that our model achieves near-perfect classification performance (macro F1-score ≈ 99.55%, Subset Accuracy ≈ 98.43%) on the test set, significantly outperforming several state-of-the-art transfer-learning models (e.g., EfficientNet, ResNet, DenseNet) in both accuracy and efficiency. Notably, CDLA has a modest model size (~10.9 MB) and fast inference time (~9.3 ms per image), making it deployable on mobile and embedded platforms. These findings indicate that the proposed architecture can effectively detect multiple concurrent leaf diseases in real time, offering a practical tool for precision agriculture.