HITL-CNN: Human-in-the-Loop Crop Disease Diagnosis Model
An EfficientNet-B4-based convolutional neural network for crop disease
classification, combined with Monte Carlo Dropout uncertainty
quantification and a confidence-threshold-based Human-in-the-Loop (HITL)
referral mechanism. Developed as part of an MRes dissertation at the
University of Greater Manchester.
Live demo: hitl-crop-diagnosis.streamlit.app
Code: https://github.com/INITREASURE0611/hitl-cnn-app
Model Description
- Architecture: EfficientNet-B4, ImageNet-pretrained backbone
- Input: 224×224 RGB leaf images
- Output: 38-class disease/healthy classification across 14 crop species
- Uncertainty quantification: Monte Carlo Dropout (T=50 stochastic
passes, p=0.1), with post-hoc temperature scaling for calibration
- Referral mechanism: predictions with entropy above a calibrated
threshold (θ=0.45 bits) are flagged for expert review rather than
auto-classified
Training Data
Trained on the
PlantVillage dataset
(Hughes and Salathé, 2015), 54,305 images across 38 classes, licensed
CC BY-NC-SA 4.0. Split 70/15/15 (train/val/test), stratified, seed=42.
Training Procedure
Two-stage fine-tuning: Stage 1 trains the classification head only
(Adam, lr=1e-3, 2 epochs); Stage 2 fine-tunes the full network (Adam,
lr=1e-4, ReduceLROnPlateau scheduler, early stopping patience=5).
Training converged at epoch 19 (best checkpoint epoch 14).
Evaluation Results
Evaluated on a held-out test set of 8,146 images.
| Metric | Value |
|---|
| Accuracy | 99.69% |
| Macro F1 | 0.9943 |
| Macro Precision | 0.9946 |
| Macro Recall | 0.9941 |
| AUC-ROC (macro) | 1.0000 |
| ECE (post-calibration) | 0.0025 |
| HITL Referral Rate | 2.6% |
| Error Capture Rate | 88.0% |
A controlled, protocol-matched comparison against a self-trained
ResNet-50 baseline found ResNet-50 achieved marginally but statistically
significantly higher accuracy (99.84% vs. 99.69%, McNemar's p=0.0227).
This model's contribution is not claimed to rest on superior raw
accuracy, but on the combination of near-ceiling accuracy with the
calibrated HITL referral mechanism described above.
Limitations
- Trained and evaluated exclusively on laboratory-condition images;
no field-condition accuracy figure has been obtained
- The Monte Carlo Dropout uncertainty pass (T=50) requires
approximately 10 seconds per prediction on CPU, a real constraint
for real-time or low-connectivity
