🌾 Rice Disease AI – Youth Digital Citizen Challenge Submission

1. Challenge Context & Problem Statement

This project is developed for the Youth Digital Citizen Challenge: AI for Climate Changes – Twin Transition in Action.

Climate change is increasing the frequency and severity of rice diseases through altered rainfall patterns, rising temperatures, and extreme weather events. Rice farmers often detect diseases too late, leading to rapid outbreaks, yield loss, and excessive pesticide usage. There is a lack of early-warning systems that combine on-field visual evidence with spatio-temporal outbreak forecasting.

This project proposes an AI-powered decision support system that:

Detects rice leaf diseases from images taken by farmers
Filters out non-rice images to avoid noise
Forecasts the risk of disease outbreak within the next 7 days at a regional (grid) level

The system is designed for real-world deployment in hackathon conditions: lightweight models, modular APIs, and extensible pipelines.

2. Solution Overview

The solution consists of three core AI components, each implemented as an independent script/module:

Rice / Non-Rice Image Filtering
→ Ensures only valid rice images are processed downstream
Rice Leaf Disease Classification
→ Identifies the disease type from rice leaf images
7-Day Disease Outbreak Forecasting
→ Predicts whether a disease outbreak will occur in the next 7 days based on crowd-sourced reports and weather signals

These components can be deployed independently or integrated into a single backend service.

Pipeline Diagram

Input Image
   ↓
MobileNetV3-Small
(Rice / Non-Rice)
   ↓
DINOv2 + MLP
(Disease Class)
   ↓
XGBoost
(Outbreak Risk T+7d)

Model Table

| Stage | Model             | Input               | Output             | Why                        |
| ----- | ----------------- | ------------------- | ------------------ | -------------------------- |
| 1     | MobileNetV3-Small | Image               | Rice / Non-rice    | Remove noisy farmer images |
| 2     | DINOv2 + MLP      | Rice image          | Disease class      | Strong generalization      |
| 3     | XGBoost           | Tabular time-series | Outbreak risk (7d) | Non-linear interactions    |

3. System Architecture

Farmer Mobile App
      │
      ▼
[ Rice / Non-Rice Filter ]  (MobileNetV3)
      │
      ▼
[ Disease Classifier ]      (DINOv2 + MLP Head)
      │
      ▼
[ Outbreak Forecasting ]    (XGBoost – 7 days ahead)
      │
      ▼
Risk Map • Alerts • Recommendations

4. Model 1 – Rice / Non-Rice Image Filtering

Objective

Design a lightweight filtering model to remove non-relevant images (soil, grass, cattle, background scenes) before downstream rice disease classification, ensuring clean and reliable inputs.

Dataset Construction

Rice images: Rice Leaf Disease datasets

Non-rice images: Soil, grass, and cattle datasets collected from Kaggle

Data integrity:

Duplicate images across sources were removed using MD5 hashing
This prevents train–validation leakage and artificially inflated performance

Class balance strategy:

Non-rice samples were capped to limit class dominance

Model Design

Backbone: MobileNetV3-Small (ImageNet pretrained)

Task: Binary classification (Rice vs Non-Rice)

Design rationale:

Chosen for its low computational cost and suitability for edge / cloud-GPU deployment
Optimized for early-stage filtering rather than fine-grained recognition

Training Strategy

Loss: BCEWithLogitsLoss with class weighting to mitigate imbalance

Optimizer: Adam

Validation threshold: 0.8, intentionally increased to reduce false positives

Regularization: Early stopping based on validation loss

Augmentation: Spatial and color perturbations to improve robustness

Output Artifact

best_model.pt

Evaluation Results

| Metric                 | Value                                       |
| ---------------------- | ------------------------------------------- |
| Task                   | Binary detection (Rice plant vs background) |
| Model                  | MobileNetV3-Small                           |
| Training epochs        | 5 (early stopping)                          |
| Training accuracy      | 98.4% → 100%                                |
| Validation accuracy    | 88.9% – 91.1%                               |
| Precision (P)          | 1.000                                       |
| Recall (R)             | 0.42 – 0.54                                 |
| False Positive Rate    | 0.000                                       |
| Validation loss (best) | 0.0618                                      |
| Overfitting observed   | Yes                                         |
| Detection behavior     | Conservative (precision-oriented)           |

MobileNetV3-small Loss Curve

MobileNetV3-small Accuracy Curve

MobileNetV3-Small was selected as a lightweight detector for rice plant localization. While achieving very high precision (P = 1.00) and zero false positives, the model exhibited relatively low recall (R ≈ 0.45–0.54), indicating missed detections. This conservative behavior ensures clean inputs for downstream disease classification but limits overall pipeline recall. The trade-off highlights the challenge of balancing efficiency and coverage under constrained data and compute budgets.

5. Model 2 – Rice Leaf Disease Classification

Objective

Classify rice leaf disease types from cropped rice leaf images, serving as the core diagnostic component of the system.

Dataset

Source: Mendeley – Rice Leaf Disease Dataset

Split strategy:

Train / Validation / Test = 70% / 15% / 15%
Fixed random seed to ensure reproducibility

Classes: Multiple rice leaf disease categories + healthy class

Preprocessing:

Standard resizing and normalization
Augmentation applied only during training to improve robustness

Model Architecture

Backbone: DINOv2 ViT-B/14 (foundation vision model)

ImageNet-pretrained
Fully frozen during training
Used strictly as a feature extractor

Feature representation:

CLS token embedding (768-dim)
L2-normalized to stabilize classifier training

Classification head:

MLP: 768 → 256 → Num_Classes
ReLU activation + Dropout (0.3)

Why DINOv2?

Strong representation learning on unseen visual patterns

Robust to variations in lighting, background, and leaf orientation

Well-suited for low-data agricultural domains

Enables fast convergence and stable training when used as a frozen backbone

Training Strategy

Backbone frozen to prevent overfitting and reduce compute cost

Optimizer: AdamW

Loss: Cross-Entropy Loss

Regularization:

Dropout in classification head
Early stopping based on validation loss

Training paradigm: Feature extraction + lightweight supervised head

Output Artifact

dino_classifier.pt

Contains:

Trained classifier weights
Class labels
Backbone metadata

Evaluation Results

Training and Validation Loss

The model shows stable and consistent learning behavior across training epochs, with both training and validation loss decreasing steadily and no signs of divergence.

Validation accuracy improves significantly from 0.37 → 0.80, indicating that the model is able to generalize well rather than overfitting to the training data.

The small gap between training and validation accuracy in later epochs suggests effective regularization and appropriate model capacity for the dataset.

Performance gains are gradual and sustained, which is typical for multi-class image classification tasks involving visually similar disease patterns.

The model converges smoothly, with the best validation accuracy achieved at epoch 29 (80.4%), showing that extended training continues to provide meaningful improvements.

Early epochs exhibit low accuracy, reflecting the complexity of rice disease classification and high inter-class visual similarity.

Overall, the results indicate a reliable baseline model suitable for deployment or further improvement via data augmentation, fine-grained labels, or stronger backbone architectures.

6. Model 3 – 7-Day Disease Outbreak Forecasting

Objective

Forecast whether a disease outbreak will occur 7 days in the future for each spatial grid, enabling early warning and proactive intervention.

Label Construction

Outbreak labels are generated using time-shifted targets within each grid_id:

   outbreak_future_7d = outbreak(t + 7)

Temporal ordering is strictly preserved to prevent information leakage.
Samples without future observations are removed.

Feature Engineering

The model leverages short- and mid-term temporal signals combined with environmental and epidemiological context:

Disease dynamics

reports_3d, reports_7d, reports_14d
same_disease_ratio
dominant_disease (label-encoded)

User activity

unique_users_7d

Environmental factors

rainfall_7d
avg_temp_7d

Data Splitting Strategy

Time-aware split based on the 80th percentile of the date distribution

Ensures realistic forecasting by training strictly on past data and evaluating on future periods

Model Design

Algorithm: XGBoost Classifier

Chosen for:

Strong performance on tabular data
Ability to model non-linear relationships and feature interactions
Robustness to mixed-scale features and missing patterns

Output Artifacts

outbreak_forecasting_artifact.pkl
dominant_disease_encoder.pkl

Evaluation Results

Threshold	Accuracy	Precision (Outbreak)	Recall (Outbreak)	F1-score (Outbreak)
0.3	0.95	0.64	0.85	0.73
0.4	0.95	0.67	0.80	0.73
0.5	0.94	0.68	0.69	0.69
0.6	0.95	0.73	0.62	0.67

The dataset is highly imbalanced, with outbreak cases accounting for only ~9% of samples.
I evaluated the XGBoost model under multiple probability thresholds to analyze the trade-off between precision and recall for outbreak detection.
Lower thresholds (0.3–0.4) significantly improve recall, allowing the system to detect more potential outbreaks at the cost of increased false positives.
Threshold = 0.3 achieves the highest recall (0.85) and best F1-score (0.73) for the outbreak class, making it suitable for early warning systems, where missing an outbreak is more critical than false alarms.
Higher thresholds (0.5–0.6) increase precision but reduce recall, which may be preferable in scenarios with limited intervention resources.
The final threshold can be adjusted depending on risk tolerance and operational constraints.

7. Key Innovations

Foundation model (DINOv2) applied to agriculture
Multi-stage AI pipeline (filter → classify → forecast)
Temporal forecasting instead of static prediction
Designed for real farmer-generated data

8. Alignment with Twin Transition (Green × Digital)

This project directly supports the Twin Transition vision:

Green transition: early disease detection reduces crop loss and limits excessive chemical usage, improving environmental sustainability
Digital transition: AI, foundation models, and data-driven forecasting empower farmers with real-time decision support

By integrating AI into climate-resilient agriculture, the solution demonstrates how digital technologies can accelerate green outcomes.

9. Potential Extensions

Disease severity estimation
Treatment & pesticide recommendation engine
Real-time outbreak heatmap
Farmer incentive system (points, vouchers)

10. Tech Stack

PyTorch, TorchVision
DINOv2 (Meta AI)
XGBoost
Scikit-learn

11. Reproducibility – How to Run on Google Colab (GPU)

To ensure transparent, reproducible, and accessible evaluation, this project is designed to be run entirely on Google Colab using a T4 GPU.

11.1 Environment Setup

Open Google Colab

Select Runtime → Change runtime type

Set:

Hardware accelerator: GPU
Recommended GPU: T4

Verify GPU availability:

import torch
print(torch.cuda.get_device_name(0))

11.2 Install Dependencies

!pip install -q torch torchvision torchaudio
!pip install -q xgboost scikit-learn pandas seaborn matplotlib kagglehub joblib

11.3 Dataset Preparation

Rice Leaf Disease Dataset

!wget -O rice_leaf.zip "https://data.mendeley.com/public-files/datasets/hx6f852hw4/files/bb4d2847-1cab-4e56-aa7a-4aa73f6dcdfc/file_downloaded"
!unzip rice_leaf.zip -d /content/dataset

External Datasets (Non-rice filtering)

These datasets are automatically downloaded via kagglehub inside the notebook.

⚠️ Note: Kaggle authentication is required when running for the first time.

11.4 Run Order (Recommended)

Rice / Non-Rice Filtering Model → Build clean input pipeline

Rice Leaf Disease Classification → Train DINOv2-based classifier

7-Day Outbreak Forecasting → Train XGBoost temporal prediction model

Each script is independent and can be executed sequentially in Colab.

11.5 Expected Artifacts

After successful execution, the following artifacts will be generated:

best_model.pt                       # Rice / Non-rice filter
dino_classifier.pt                  # Disease classifier
outbreak_forecasting_artifact.pkl   # Forecasting model
dominant_disease_encoder.pkl        # Label encoder

These artifacts can be reused directly for deployment or API integration.

11.6 Reproducibility Notes

Fixed random seeds for dataset splitting

Frozen backbone for disease classification

Clear threshold definitions for forecasting & filtering

This setup ensures that any evaluator or participant can reproduce the results by simply uploading the notebook to Colab and running all cells.

12. Team Notes (Youth Digital Citizen Challenge)

This project is designed to be:

Modular – each model can be deployed independently
Scalable – supports region-level outbreak monitoring
Practical – aligned with real farming workflows

✨ Empowering farmers with early, AI-driven disease intelligence.