Detecting Plant Diseases with Vision Transformers

Urban Farming Meets Deep Learning

Urban farming is growing rapidly, but plant diseases remain a major threat to crop yields. Early detection is critical — but most urban farmers lack the expertise to identify diseases at early stages. I built a computer vision system that automates this process using deep learning.

CNN vs. Vision Transformer

The project compared two architectures:

Convolutional Neural Networks (CNN)

CNNs have been the standard for image classification for years. They use local receptive fields to detect patterns like edges, textures, and shapes hierarchically.

import torch.nn as nn

class PlantCNN(nn.Module):
    def __init__(self, num_classes):
        super().__init__()
        self.features = nn.Sequential(
            nn.Conv2d(3, 64, kernel_size=3, padding=1),
            nn.BatchNorm2d(64),
            nn.ReLU(inplace=True),
            nn.MaxPool2d(2),
            # ... deeper layers
        )
        self.classifier = nn.Linear(512, num_classes)

Vision Transformers (ViT)

Vision Transformers split the image into patches and process them as a sequence — similar to how language models process words. This gives them a global view of the image from the very first layer.

The ViT model showed better performance on our dataset, particularly for diseases that manifest as subtle, distributed patterns across the leaf surface.

Making It Explainable with GradCAM

A model that says "this plant is diseased" isn't enough. Farmers need to understand why and where. That's where GradCAM (Gradient-weighted Class Activation Mapping) comes in.

GradCAM generates a heatmap showing which regions of the image most influenced the model's prediction. For plant disease detection, this highlights the exact areas of the leaf showing disease symptoms.

from pytorch_grad_cam import GradCAM
from pytorch_grad_cam.utils.image import show_cam_on_image

cam = GradCAM(model=model, target_layers=[model.layer4[-1]])
grayscale_cam = cam(input_tensor=input_tensor)

# Overlay heatmap on original image
visualization = show_cam_on_image(rgb_img, grayscale_cam[0], use_rgb=True)

This explainability layer was crucial for building trust with farmers who are understandably skeptical of AI making decisions about their crops.

Key Results

The Vision Transformer achieved higher accuracy but at the cost of slower inference. For a real-time mobile application, the CNN might be preferred; for batch processing of captured images, the ViT is the clear winner.

What I Learned

• Data quality trumps model architecture — Careful data augmentation and cleaning improved both models more than any architecture change.
• Explainability drives adoption — GradCAM wasn't just a nice-to-have; it was the feature that made farmers trust the system.
• Domain expertise is essential — Working with plant pathologists to validate model predictions caught errors that pure metrics would have missed.

Computer vision in agriculture is a field where AI can make a tangible, real-world impact. The technology exists — the challenge is making it accessible and trustworthy for the people who need it most.