adaptive-edge-plant-model

MobileNetV4 Edge Plant Classifier (with EDL & GRL)

This repository hosts the lightweight Edge Classifier (MobileNetV4-Conv-Medium) for the Adaptive Edge-Cloud Plant Disease Diagnosis framework. The model features:

1. Evidential Deep Learning (EDL) Head to calculate epistemic uncertainty (vacuity $u$) in a single forward pass.
2. Gradient Reversal Layer (GRL) Domain Adaptation to reconcile lab-to-field domain shifts.
3. Conformal Temperature Scaling configuration parameters to provide distribution-free confidence guarantees.

1. Mathematical and Framework Documentation

A complete mathematical report detailing the Evidential Deep Learning (EDL), Unsupervised Domain Adaptation (UDA), and Conformal Calibration models is compiled as a PDF and available in this repository: 👉 Read the Mathematical Report (PDF)

---

2. Model Architecture and Training Details

• Model Type: MobileNetV4-Conv-Medium with Evidential Classification Head & Domain Discriminator
• Number of Classes: 88 (spanning various crop types including Apple, Tomato, Wheat, Soybean, Sugarcane, Tea, etc.)
• Optimization Strategy: Trained from scratch with all layers unfrozen to adapt features to the plant classification domain.
• Optimizer: AdamW (Learning Rate: $10^{-3}$, Weight Decay: $10^{-3}$)
• Loss Function: Multi-Task Evidential Loss ($\mathcal{L}{mse} + \lambda_t \mathcal{L}{kl}$) with auxiliary Cross-Entropy ($\gamma = 0.1$) to prevent gradient vanishing.
• Domain Adaptation: Unsupervised Domain Adaptation (UDA) with Gradient Reversal Layer (GRL) mapping source (lab) to target (field) domains.

Training & Validation Loss Curves

Accuracy & Validation Vacuity Dynamics

---

3. Convergence Metrics Summary

The model was trained for 10 epochs on Kaggle GPU environments using a stratified dataset split (30,103 training images, 5,347 validation images):

Note: The best-performing checkpoint was recorded at Epoch 9 with 92.23% validation accuracy and the lowest average evidential uncertainty (vacuity = 0.8219).

---

4. Collaborative Gating Mechanism

The Edge classifier is designed to run locally on resource-constrained devices. It makes predictions and computes the epistemic uncertainty (vacuity $u$) in a single forward pass.

• If vacuity exceeds the threshold ($u > au_{vac}$) OR maximum calibrated conformal confidence is below the threshold ($p_{max} < au_{conf}$), the diagnostic request is offloaded to the heavy cloud model (ConvNeXt-Large).