Plant Disease Detection Using Mobile Images and Video

Pixel-Level Semantic Segmentation

Bounding boxes give rectangular regions, but plant disease spots have irregular, curved shapes. Semantic segmentation classifies every pixel in the image individually: is this pixel healthy leaf tissue, a diseased lesion, or background [3], [4]?

Common Segmentation Models

• U-Net: Uses an encoder-decoder architecture with skip connections that pass fine detail from the downsampling path to the upsampling path, helping preserve precise lesion boundaries.
• DeepLabV3+: Uses Atrous Spatial Pyramid Pooling (ASPP) to capture patterns at multiple scales without losing resolution, making it effective for complex leaf and spot outlines [4], [5].

Disease Severity Estimation

In farming, knowing that a leaf has a disease is only part of the story. A leaf with 2% damage might only need monitoring, whereas 40% damage might require immediate treatment. With segmentation, we can calculate the exact severity percentage:

Algorithm 1: Disease Severity Calculation

Once the segmentation model outputs the pixel counts for the diseased area and the total leaf, the severity percentage is calculated:

# Disease Severity Estimation Algorithm
# Inputs: Segmented pixel counts from DeepLabV3+ or U-Net

def calculate_disease_severity(diseased_pixels: int, total_leaf_pixels: int) -> float:
    """Calculates the percentage of leaf area affected by disease."""
    if total_leaf_pixels <= 0:
        raise ValueError("Total leaf area must be greater than zero.")
    
    severity_percentage = (diseased_pixels / total_leaf_pixels) * 100.0
    return severity_percentage

# Example with segmented pixel counts:
diseased_pixels = 2340
total_leaf_pixels = 10000

severity = calculate_disease_severity(diseased_pixels, total_leaf_pixels)
print(f"Disease Severity: {severity:.2f}%")

Algorithm Steps

1. Segment the complete leaf from the background.
2. Segment the diseased spots (chlorosis or necrosis).
3. Count total leaf pixels and diseased pixels.
4. Divide diseased pixels by total leaf pixels and multiply by 100.