KrishokChat: Agricultural Vision & Soil Intelligence Model Suite

This repository contains the complete, production-ready suite of machine learning models developed for the KrishokChat Advisory System. These models are packaged into self-contained, folderized directories to enable academic reviewers, agronomists, and developers to reproduce and run our full diagnostic pipeline locally without requiring expensive cloud GPU servers.

1. Architectural Overview

The KrishokChat perception and diagnostic stack uses a multi-tier decoupled architecture:

1. Tier 1: Crop Species Identification — A YOLO26-cls model identifying 10 major crop families in Bangladesh to eliminate cross-crop pesticide misbinding hazard.
2. Tier 2: Specialized Crop Pathology Diagnosis — Six per-crop dedicated YOLO26-cls models diagnosing foliar fungal, bacterial, viral, and physiological diseases.
3. Tier 3: Soil Moisture Matric Tension Estimation — A 5-fold cross-validated EfficientNet-B0 regression model predicting continuous soil suction pressure ($\text{kPa}$) directly from field images for precision irrigation guidance.

                   [ Farmer Field Leaf Image ]
                               │
                               ▼
                   [ 01_crop_classifier ] (YOLO26-cls)
             ┌─────────────────┼─────────────────┐
             │                 │                 │
             ▼                 ▼                 ▼
     [ 02_potato_disease ]  [ 03_rice_disease ]  [ 04_wheat_disease ] ...
             │                 │                 │
             └─────────────────┼─────────────────┘
                               ▼
               [ Diagnosis + Bangla Treatment Advice ]

2. Model Catalog & Checkpoint Manifest

Every PyTorch weight (.pt) is named distinctively with zero ambiguity. Corresponding optimized ONNX weights are included for cross-platform edge deployment.

3. Directory Layout

backend_models/
├── README.md                           # Main Model Card & Reviewer Documentation
├── requirements.txt                    # Minimal pip dependencies
├── verify_all_models.py                # Single-click automated diagnostic verification suite
├── run_inference.py                    # Multi-model CLI inference runner
├── upload_to_huggingface.py            # Hugging Face Hub upload automation script
│
├── 01_crop_classifier/                 # Intake Crop Classifier (10 classes)
│   ├── crop_classifier.pt
│   ├── crop_classifier.onnx
│   ├── class_names.json
│   └── README.md
│
├── 02_potato_disease/                  # Potato Pathology (Early Blight, Late Blight, Healthy)
│   ├── potato_disease.pt
│   ├── potato_disease.onnx
│   ├── class_names.json
│   ├── disease_details.json
│   └── README.md
│
├── 03_rice_disease/                    # Rice Pathology (Blast, BLB, Sheath Blight, etc.)
│   ├── rice_disease.pt
│   ├── rice_disease.onnx
│   ├── class_names.json
│   ├── disease_details.json
│   └── README.md
│
├── 04_wheat_disease/                   # Wheat Pathology (Rusts, Blast, Tan Spot, etc.)
│   ├── wheat_disease.pt
│   ├── wheat_disease.onnx
│   ├── class_names.json
│   ├── disease_details.json
│   ├── metadata.json
│   └── README.md
│
├── 05_corn_disease/                    # Corn Pathology (Rust, Leaf Spot, Blight)
│   ├── corn_disease.pt
│   ├── corn_disease.onnx
│   ├── class_names.json
│   ├── disease_details.json
│   └── README.md
│
├── 06_chilli_disease/                  # Chilli Pathology (Bacterial Spot, Curl Virus, etc.)
│   ├── chilli_disease.pt
│   ├── chilli_disease.onnx
│   ├── class_names.json
│   ├── disease_details.json
│   └── README.md
│
├── 07_brassica_disease/                # Crucifers (Cabbage / Cauliflower diseases)
│   ├── brassica_disease.pt
│   ├── brassica_disease.onnx
│   ├── class_names.json
│   ├── disease_details.json
│   └── README.md
│
├── 08_soil_moisture/                   # Soil Moisture Tensiometer Regression (5-fold ensemble)
│   ├── effnetb0_fold0.pt ... fold4.pt
│   ├── oof_predictions.csv
│   ├── pred_vs_actual.png
│   ├── model_def.py
│   └── README.md
│
└── sample_images/                      # Representative leaf images for zero-setup verification
    ├── crop_sample.jpg
    ├── potato_sample.jpg
    ├── rice_sample.jpg
    ├── wheat_sample.jpg
    ├── corn_sample.jpg
    ├── chilli_sample.jpg
    └── brassica_sample.jpg

4. Quickstart Guide for Academic Reviewers

Step 1: Install Dependencies

cd backend_models
pip install -r requirements.txt

Step 2: Automated Verification of All Models

Run the automated test runner to verify that every single model checkpoint loads into memory, performs forward inference, and verifies against class specifications:

python verify_all_models.py

Step 3: Run Standalone Inferences

Test Potato Disease Diagnosis:

python run_inference.py --crop potato --image sample_images/potato_sample.jpg

Test Rice Disease Diagnosis:

python run_inference.py --crop rice --image sample_images/rice_sample.jpg

Test Wheat Disease Diagnosis:

python run_inference.py --crop wheat --image sample_images/wheat_sample.jpg

Test Soil Moisture Regression:

python run_inference.py --task soil --image sample_images/crop_sample.jpg

5. Python API Usage

Vision Models (YOLO26)

from ultralytics import YOLO

# Load any specialized YOLO26 pathology classifier
model = YOLO("02_potato_disease/potato_disease.pt")

# Predict on an RGB image
results = model.predict("sample_images/potato_sample.jpg", imgsz=224, verbose=False)[0]

top_class = model.names[results.probs.top1]
confidence = float(results.probs.top1conf)
print(f"Result: {top_class} ({confidence:.2%})")

Soil Moisture Regression (5-Fold Ensemble)

import sys
sys.path.insert(0, "08_soil_moisture")
from model_def import load_soil_model
import torch
import torchvision.transforms as T
from PIL import Image

transform = T.Compose([
    T.Resize((224, 224)),
    T.ToTensor(),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

img = transform(Image.open("sample_images/crop_sample.jpg").convert("RGB")).unsqueeze(0)

# Ensemble average across 5 folds
predictions = []
for fold in range(5):
    m = load_soil_model(f"08_soil_moisture/effnetb0_fold{fold}.pt")
    with torch.no_grad():
        predictions.append(m(img).item())

mean_kpa = sum(predictions) / len(predictions)
print(f"Estimated Soil Matric Potential: {mean_kpa:.2f} kPa")

6. Uploading to Hugging Face Hub

To publish this suite to the official Hugging Face repository:

python upload_to_huggingface.py --token <YOUR_HF_TOKEN>

Target repository: https://huggingface.co/RaiyanKhaan/KrishokTech-Models

7. Citation

If you use these model weights or the KrishokChat pipeline in your research, please cite our paper:

@inproceedings{krishokchat2026,
  title={KrishokChat: A Grounded Multi-Modal Agricultural Advisory and Pathology Diagnostic System for Smallholder Farming},
  author={Reza, Raiyaan and Khaan, Raiyan and Collaborators},
  booktitle={Proceedings of the European Chapter of the Association for Computational Linguistics (EACL)},
  year={2026}
}