Abstract:
Early and accurate diagnosis of mango (Mangifera indica) leaf diseases is
essential to reduce yield losses and avoid unnecessary pesticide use, yet reliable
field-ready tools remain limited. This study evaluates whether a stacked deeplearning ensemble can deliver robust, generalizable classification of visually
similar leaf pathologies while remaining deployable on mobile hardware. A
curated image corpus of 6,400 samples spanning eight classes (seven diseases
and healthy) was standardized and partitioned into stratified training,
validation, and test sets. Five pretrained convolutional networks (VGG16,
ResNet50, InceptionV3, DenseNet121, Xception) were fine-tuned under identical
preprocessing and regularization regimes; their calibrated softmax outputs were
concatenated and learned by a VGG16 meta-classifier. Performance on the heldout test set was assessed using overall accuracy and macro-averaged precision,
recall, and F1, with class-wise confusion matrices and ROC–AUC to characterize
separability. The ensemble achieved 99.38% accuracy with macro precision,
recall, and F1 ≈ 0.99, outperforming the best single models (DenseNet121 and
VGG16) and reducing confusions among symptomatically related classes. GradCAM analyses indicated that predictions were driven by lesion margins and
chlorotic patterns rather than background artifacts, supporting diagnostic
plausibility. The trained model was converted to TensorFlow Lite for on-device
inference, demonstrating sub-second latency on commodity mobile hardware
without network connectivity. These findings provide evidence that stacked
ensembling of strong CNN backbones yields near-ceiling diagnostic performance
for mango leaf diseases while satisfying practical constraints for field
deployment, with potential to accelerate agronomic decision-making and
promote more judicious chemical use in smallholder systems.
