language:
- en
license: mit
tags:
- image-classification
- pytorch
- convnext
- plant-disease
- agriculture
- edge-cloud
pipeline_tag: image-classification
model-index:
- name: ConvNeXt-Large Cloud Plant Classifier
results:
- task:
type: image-classification
name: Image Classification
dataset:
name: Plant Disease Classification Merged Dataset (88 Classes)
type: plant-disease-classification-merged-dataset
metrics:
- type: accuracy
value: 96.42
name: Validation Accuracy
ConvNeXt-Large Cloud Plant Classifier
This repository hosts the ConvNeXt-Large cloud classifier model trained as part of the Confidence-Aware Adaptive Edge–Cloud Framework for Reliable Plant Disease Diagnosis in Low-Connectivity Agricultural Environments.
The cloud model serves as the high-capacity, heavy diagnostician tier of the cooperative inference pipeline. It is triggered dynamically by the edge node when a captured leaf image exhibits high epistemic uncertainty (vacuity $u > au_{vac}$) or low conformal confidence ($p_{max} < au_{conf}$).
Model Architecture
- Backbone: ConvNeXt-Large (originally pre-trained on ImageNet-1K)
- Classifier Head: Modified Linear Layer projecting to 88 pathological crop classes
- Input Dimension: $384 imes 384$ pixels (higher resolution to capture fine-grained leaf spots)
Training Setup & Parameters
- Dataset: Merged Plant Disease Dataset (88 unique classes spanning laboratory and field-captured folders).
- Optimization Strategy: Parameter-Efficient Fine-Tuning (PEFT) with convolutional stages 1–3 frozen, training only the 4th stage and classifier head.
- Optimizer: AdamW (Learning Rate: 1e-4, Weight Decay: 1e-3)
- Loss Function: Cross-Entropy Loss
- Train/Val Split: 85% Train / 15% Validation (Stratified per-folder split)
Operational Performance (Epoch 7 Results)
Below are the recorded metrics from the training session:
| Metric | Score / Value | Description |
|---|---|---|
| Training Loss | 0.1039 |
Cross-Entropy loss at epoch 7 |
| Training Accuracy | 96.47% |
Accuracy over the training split |
| Validation Loss | 0.1028 |
Cross-Entropy loss on the held-out split |
| Validation Accuracy | 96.42% |
Calibrated validation classification rate |
Integration & Deployment
The model checkpoint is designed to be hosted in a FastAPI-based Docker container on Hugging Face Spaces (Port 7860), responding to offloading queries via HTTP REST or binary JSON over MQTT v5 brokers.