Ensemble Deep Learning for Aloe Vera Leaf Disease Classification

Aloe Vera is a widely cultivated medicinal plant, yet automated disease diagnosis for it remains comparatively underexplored relative to major food crops. This study presents a rigorous evaluation of deep learning approaches for classifying Aloe Vera leaves as healthy, rot-affected, or rust-affected, drawn from the same public repository used to benchmark a recent lightweight architecture, AloeVeraNet, which reported 96.09% accuracy. We fine-tune and fairly compare four ImageNet-pretrained backbones (MobileNetV2, EfficientNet-B0, ResNet18, EfficientNetB3) under an identical stratified evaluation protocol, finding that backbone choice has limited impact among compact architectures (97.14%–97.71% accuracy). Motivated by this finding and by confusion-matrix evidence that rot/rust discrimination is the dominant source of error, we investigate a two-stage hierarchical classification pipeline. Contrary to results reported for this architecture pattern in a related plant-disease detection task, we find it underperforms flat classification (94.10% vs. 97.71%). A simple ensemble of the four flat models instead achieves the best result, 98.86% accuracy, surpassing this previously reported accuracy, though this improvement does not reach conventional statistical significance (McNemar's test, p = 0.109). Grad-CAM visualizations further confirm the model attends to biologically relevant lesion regions.