🌿 CropSense: Quality Detection in Solanaceous Crops Using Deep Learning

A novel multi-headed CNN architecture for simultaneous classification of potato and tomato crops with quality assessment using explainable AI.

📋 Table of Contents

• Overview
• Key Features
• Dataset
• Installation
• Usage
• Results
• Model Explainability
• Deployment
• Future Work
• Contributors

🎯 Overview

CropSense is a deep learning-based system that performs dual-task classification:

1. Crop Classification: Distinguishes between potato and tomato crops
2. Quality Assessment: Identifies healthy vs. diseased samples

Traditional agricultural inspection methods are time-consuming and error-prone. This project bridges the gap between crop identification and quality control using a unified multi-task learning framework with 99.9% crop classification accuracy and 98.5% quality assessment accuracy[attached_file:1].

Problem Statement

Manual crop inspection by experts is:

• Time-consuming and labor-intensive
• Prone to human error
• Not scalable for large-scale agriculture

Most existing automated solutions focus on either classification OR quality assessment in isolation, missing opportunities for shared feature learning[attached_file:1].

✨ Key Features

• Multi-Task Learning: Shared feature extractor with task-specific heads
• Lightweight Architecture: 30% parameter reduction compared to separate models
• Explainable AI: Grad-CAM visualizations for transparent decision-making
• High Performance:
  • 99.9% accuracy in crop classification (potato vs. tomato)
  • 98.5% accuracy in quality assessment (healthy vs. diseased)
• Production-Ready: Streamlit web interface for real-time predictions
• Robust Data Augmentation: Geometric and photometric transformations
• Fast Training: Optimized for NVIDIA P100 GPU (26 minutes/50 epochs)

Optimization Framework

• Loss Function: Unweighted sum of cross-entropy losses
• Optimizer: Adam with cyclic learning rate (1e-5 to 1e-3)
• Regularization:
• L2 weight decay (λ = 0.01)
• 50% dropout
• Batch normalization
• Training Protocol:
• Early stopping (patience=7)
• Learning rate reduction on plateau (factor=0.2, patience=2)
• Max 50 epochs, batch size 32

📊 Dataset

Dataset Composition

Custom balanced dataset of 10,000 images (256×256 resolution):

Category	Healthy	Diseased	Total
Potato	2,500	2,500	5,000
Tomato	2,500	2,500	5,000
Total	5,000	5,000	10,000

Data Sources

1. Tomato Dataset (IEEE Access): 7,226 images across 4 states (Unripe, Ripe, Old, Damaged)

• Old/Damaged → Diseased
• Unripe/Ripe → Healthy

2. Potato Dataset: 200,000+ images covering diseases like Common Scab, Dry Rot, Gangrene, Violet Root Rot[attached_file:1]

Data Split

• Training Set: 8,000 images (80%)
• Testing Set: 2,000 images (20%)

Augmentation Techniques

• Rotation
• Horizontal/Vertical flipping
• Brightness adjustments
• Zoom/Shift transformations

🚀 Installation

Prerequisites

• Python 3.8+
• CUDA 11.0+ (for GPU acceleration)

Clone Repository

git clone "https://github.com/arpanpramanik2003/cropsense.git" cd cropsense

Install Dependencies

• pip install -r requirements.txt

Output:

• Crop Prediction: tomato (99.91% confidence)
• Condition Prediction: Disease (99.40% confidence)

Streamlit Web Application

• streamlit run app.py

Access the interface at http://localhost:8501

Features:

• Drag-and-drop image upload
• Real-time predictions
• Grad-CAM visualizations
• Batch processing (up to 20 images/sec)
• Confidence threshold filtering (>90%)

📈 Results

Classification Performance

Crop Classification (Test Set)

Class	Precision	Recall	F1-Score	Support
Potato	1.00	1.00	1.00	1000
Tomato	1.00	1.00	1.00	1000
Accuracy			1.00	2000

Quality Assessment (Test Set)

Class	Precision	Recall	F1-Score	Support
Disease	0.99	0.98	0.98	1000
Healthy	0.98	0.99	0.99	1000
Accuracy			0.98	2000

Training Metrics

• Crop Classification Loss: 0.0037 (train), 0.0088 (validation)
• Quality Assessment Loss: 0.0457 (train), 0.0401 (validation)
• Training Time: 26 minutes on NVIDIA P100 GPU
• Convergence: Optimal at epoch 18 (crop), epoch 20 (quality)[attached_file:1]

Confusion Matrix Analysis

• Crop Classification: Only 2 misclassifications out of 2000 samples (0.1% error)
• Quality Assessment: 30 total errors (17 false healthy, 13 false diseased)[attached_file:1]

🔍 Model Explainability

Grad-CAM Visualization

We implement Gradient-weighted Class Activation Mapping (Grad-CAM) to provide visual explanations for both tasks:

Key Insights (200 validation samples)

Crop Classification:

• Potato: Focuses on skin texture and shape contours (α = 0.79 ± 0.12)
• Tomato: Highlights stem attachment and surface gloss

Quality Assessment:

• Healthy: Diffuse activations (mean σ = 0.18)
• Diseased: Localized at lesion sites (89% correlation with visible defects)[attached_file:1]

🌐 Deployment

Streamlit Interface Features

• Image Upload: Drag-and-drop or browse (JPG, JPEG, PNG)
• Real-Time Processing: Instant predictions with confidence scores
• Dual Heatmaps: Separate Grad-CAM visualizations for crop and quality tasks
• Batch Processing: Handle multiple images simultaneously
• Quality Thresholds: Filter predictions below 90% confidence

🔮 Future Work

Dataset Expansion

• Incorporate 5,000+ field-captured images with natural occlusions
• Add temporal growth stage variations
• Include multiple disease severity levels

Model Optimization

• Quantization: Target <50MB for Raspberry Pi deployment
• Acceleration: Implement TensorRT for real-time processing
• Compression: Knowledge distillation for edge devices

Application Enhancement

• Disease Severity Quantification: Mild/Moderate/Severe classification
• Multilingual Support: Regional language interfaces for farmers
• Mobile App: Android/iOS deployment
• Multi-Crop Extension: Expand to other vegetable families
• Multispectral Integration: Incorporate NIR/thermal imaging

👥 Contributors

• Arpan Pramanik - Dept. of CSE, The Neotia University
  📧 pramanikarpan089@gmail.com

• Diya Chanda - Dept. of CSE, The Neotia University
  📧 chandasujata01@gmail.com