Arko007/adaptive-cloud-plant-model

ConvNeXt-Large Cloud Plant Disease Diagnostician

This repository hosts the high-capacity Cloud Classifier (ConvNeXt-Large) for the Adaptive Edge-Cloud Plant Disease Diagnosis framework. The model is dynamically queried by edge nodes running MobileNetV4-EDL when classification confidence is low or uncertainty is high.

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: ConvNeXt-Large
• Number of Classes: 88 (spanning various crop types including Apple, Tomato, Wheat, Soybean, Sugarcane, Tea, etc.)
• Resolution: $384 imes 384$ pixels
• Optimization Strategy: Stages 1-3 of the backbone were frozen to optimize GPU efficiency.
• Optimizer: AdamW (Learning Rate: $10^{-4}$, Weight Decay: $10^{-3}$)
• Loss Function: Categorical Cross-Entropy

Training & Validation Loss Curves

Training & Validation Accuracy Curves

---

3. Convergence Metrics Summary

The model was trained for 7 epochs on Kaggle GPU environments using a stratified dataset split (30,103 training images, 5,347 validation images):

Epoch	Training Loss	Training Accuracy (%)	Validation Loss	Validation Accuracy (%)
1	0.6872	82.83%	0.1986	93.44%
2	0.2168	93.04%	0.1420	95.05%
3	0.1629	94.61%	0.1330	95.37%
4	0.1394	95.45%	0.1153	95.80%
5	0.1235	95.83%	0.1167	96.08%
6	0.1104	96.24%	0.1152	95.84%
7	0.1039	96.47%	0.1028	96.42%

Note: Checkpoints for all epochs, including the best-performing convnext_large_cloud_best.pth, are stored directly in the Hugging Face model repository.

---

4. Collaborative Gating Mechanism

The Cloud model acts as a secondary diagnostician in the cooperative pipeline. The lightweight edge node decides whether to query this model by checking:

1. Evidential Vacuity Threshold ($u > \tau_{vac}$): Triggers if the leaf image has out-of-distribution patterns or high epistemic uncertainty.
2. Conformal Confidence Threshold ($p_{max} < \tau_{conf}$): Triggers if the calibrated categorical prediction confidence is low.