1 of 13

GOVT. P.G. COLLEGE, DHARAMSHALA

DEPT. OF COMPUTER SCIENCE

B.TECH. (C.S.E) 6TH SEMESTER

Presented By : Ctrl+RRR

2 of 13

PLANT DISEASE

DETECTION

3 of 13

Context

800M+

People at risk

People at risk of food insecurity globally

$220B

Annual global loss

Annual global loss from crop diseases

78%

Crop losses

Crop losses caused by plant diseases

Global Food Security

Agriculture feeds over 8 billion people worldwide. Any disruption to crop health directly threatens global food supply chains, impacting food availability, prices, and nutritional security for billions.

Disease Impact

Plant diseases destroy up to 40% of food crops annually, causing severe economic and humanitarian consequences. These losses translate to billions in revenue lost and millions facing food insecurity.

Detection Gap

Most smallholder farmers lack access to plant pathologists and diagnostic laboratories. Diseases often go undetected until visible symptoms appear, by which point significant crop damage has already occurred.

AI as the Solution

Deep learning models detect disease from leaf images with expert-level accuracy, deployable on any smartphone. This technology democratizes access to expert diagnostics, enabling real-time intervention.

India-Specific Need

India's 140 million+ farming households need affordable, multilingual tools. Making Hindi-language AI interfaces critical for adoption and impact in rural communities.

4 of 13

Introduction

What is Plant AI?

An end-to-end AI web application that identifies plant leaf diseases from photos in seconds, combining computer vision with deep learning to provide accessible diagnostic capabilities previously available only to agricultural experts.

YOLO Detection

YOLOv8 object detection for precise leaf localization, automatically cropping the region of interest before classification to improve accuracy in cluttered backgrounds.

Dual Ensemble

MobileNetV2 contributes 55% weight while YOLOv8-cls contributes 45%, combining their strengths for robust classification across diverse conditions.

3 Input Modes

Upload photos from gallery, capture new images with camera, or use live webcam for real-time detection with bounding-box overlay.

Bilingual

English and Hindi UI with correctly-shaped Devanagari script in PDF reports, ensuring accessibility for India's diverse farming communities.

PDF Reports

Auto-generated per-image diagnostic reports with disease identification, treatment recommendations, and prevention guidance in both languages.

Live Detection

Real-time bounding-box overlay showing detected leaf regions and instant classification results as you point the camera at plants.

Supports 15 disease classes across Tomato, Potato, and Pepper crops — the most economically important vegetable crops in India.

5 of 13

Literature Review

Previous research established the foundation for AI-based plant disease detection, but each approach had limitations our system addresses through ensemble architecture and bilingual deployment.

Reference

Journal

Findings

Gap / Contribution

Mohanty et al. (2016)

Frontiers in Plant Science

MobileNet on PlantVillage: 99.35% accuracy in lab conditions; real-world accuracy much lower due to background noise

No leaf localization; background clutter degrades performance

Ferentinos (2018)

Comp. & Electronics in Agri.

Deep CNNs (AlexNet, VGG, GoogleNet) for disease detection; GoogleNet achieved 99.53% on lab images

High compute cost; unsuitable for mobile/edge deployment

Ramcharan et al. (2019)

Frontiers in Plant Science

MobileNetV2 smartphone app for cassava disease detection in Africa; proved mobile-first AI is feasible

Single crop; no ensemble or bounding-box detection pipeline

Jiang et al. (2020)

IEEE Access

YOLOv3 for real-time multi-class plant disease detection; 89.2% mAP at 10 FPS on embedded hardware

Classification accuracy limited without a dedicated classifier

Our System

YOLO detection + MobileNetV2 + YOLOv8-cls ensemble with bilingual UI and WeasyPrint PDF reports.

6 of 13

Problem Statement

How can resource-constrained farmers accurately identify plant diseases and receive actionable treatment guidance in real time — without expert agronomists, expensive labs, or language barriers?

1

Late Detection

Diseases identified visually only after 40–60% crop damage has occurred — far too late for effective intervention. By the time farmers notice visible symptoms, significant yield loss is already inevitable, and treatment options become limited and costly.

2

Expert Scarcity

Only 1 plant pathologist per 500+ farms in developing regions. Remote farmers have zero diagnostic access, forcing them to rely on guesswork or travel long distances to agricultural extension offices, delaying critical treatment decisions.

3

Language Barrier

Most AI diagnostic tools are English-only interfaces. 650 million+ Indian farmers primarily speak regional languages and cannot effectively use English-only systems, creating a significant adoption barrier despite technical capability.

4

No Actionable Output

Existing apps classify disease but don't produce structured, shareable treatment and prevention reports for field use. Farmers need clear, step-by-step guidance in their language, not just a disease name.

Our Solution

Real-time AI detection + bilingual guidance + auto PDF reports — accessible via any browser on smartphones, tablets, or computers without installation.

7 of 13

Methodology — System Pipeline

Our three-phase training strategy and dual-ensemble architecture maximize accuracy while maintaining efficient inference time suitable for web deployment.

📷 User Input (Upload / Camera / Live)

Images accepted in three formats: batch upload from file system, single capture via device camera, or continuous live detection from webcam feed

🔍 YOLOv8 Detector (yolov8n.pt)

YOLOv8 nano architecture detects leaf regions with bounding boxes, filtering out background clutter

◆ Leaf / Plant Found?

Validation check: if no leaf detected, HSV green blob fallback attempts to identify plant material

✂️ Crop Leaf Region (+ margin padding)

Extract detected leaf region with margin padding for classifier input

MobileNetV2 + YOLOv8-cls

Two parallel classifiers process cropped leaf image independently

⚖️ Ensemble Voting (55% MB + 45% YC)

Weighted average combines predictions for final diagnosis

📊 Result + PDF Report

Generate bilingual diagnostic report with treatment guidance

🏋️ Training Pipeline (3 Phases)

01

Data

PlantVillage Dataset — 15 classes with 54,305 images total

02

Preprocessing

Resize, augment, normalize images; split 80/20 train/validation

03

Model Training

Train MobileNetV2 and YOLOv8-cls separately, then combine

8 of 13

9 of 13

10 of 13

11 of 13

Result Analysis

~93%

MobileNetV2

Validation Accuracy

~91%

YOLOv8-cls

Top-1 Accuracy

~2 s

Per-image

Inference Time

15

Disease Classes

Detected

📋 Key Metrics

~92.4%

Precision

Correct positive predictions

~91.8%

Recall

Proportion of actual positives identified

~92.1%

F1-Score

Harmonic mean of precision and recall

0.18

Train Loss

Final training loss value

0.23

Val Loss

Final validation loss value

~0.5

Inference FPS

Frames per second processing

Ensemble boosts accuracy ~2% over single model — The combination of MobileNetV2 and YOLOv8-cls with weighted voting improves accuracy by approximately 2 percentage points compared to using either model alone, demonstrating the value of ensemble architecture.

12 of 13

Conclusion

Key Achievements

YOLO + MobileNetV2 ensemble achieves ~93% validation accuracy on 15 disease classes

Bounding-box leaf localization improves classifier precision in cluttered backgrounds

Bilingual UI (English + Hindi) with correctly-shaped Devanagari PDF reports

Three input modes: batch upload, camera snapshot, and real-time live detection

Auto-generated WeasyPrint PDF with diagnosis, treatment, and prevention per image

3-phase training strategy maximizes accuracy with efficient resource use

🚀 Future Enhancements

📱 Mobile App

Flutter/React Native offline field app for farmers without consistent internet connectivity

🌾 More Crops

Expand to Rice, Wheat, Maize, Sugarcane — India's major staple crops

🗺 GPS Heatmapping

Map disease spread by region for alerts and agricultural planning

☁️ Edge Deployment

Raspberry Pi + TFLite for remote farms without cloud infrastructure

🌍 More Languages

Tamil, Telugu, Marathi, Bengali support for pan-India reach

📊 Severity Scoring

Estimate disease severity percentage per image for treatment intensity

🤝 Community Portal

Crowd-sourced data to retrain models with real-world farmer-submitted images

13 of 13

Thank You