Deep Learning for Disease Resistance Classification using Drone Imagery of Corn

Project Overview

Corn breeders currently rely on slow, subjective visual scoring to assess foliage disease resistance across thousands of plots. This project reframes disease assessment as a supervised regression task: training a deep neural network to predict Southern Leaf Blight (SLB) severity directly from drone-collected RGB images. The challenge lies in designing a system that generalizes across seasons and field conditions while managing large, noisy, and imbalanced real-world data.

Data

Disease Scoring Image

The dataset is composed of approximately 10,000 images of the plants and associated foliage health scores on a scale of 1 (highly diseased) to 9 (no disease) with 0.5 size increments. Disease severity scores were collected by PhD students and researchers in plant pathology over two years.

Numerous CNN models were trained and tested to evaluate custom model architectures as well as off the shelf model implementations including EfficientNetv2 and ResNet. Custom models were implemented in PyTorch and using Optuna and MLFlow we were able to train and test numerous variations of convolutional and fully connected layers, adjusting both the depth and width of the models. Initial training struggled to learn the data effectively with accuracy remaining below 40%. Deeper evaluation of the images used for training led us to suspect that the models were having difficulty identifying and focusing on appropriate corn foliage and disease features and were not able to filter out and ignore the other non-relevant features in the images like non-corn vegetation, soil and shadows.

Corn Image Filtering

To augment our training images, we decided to segment the images to filter out the non-relevant features and allow the models train on images that were segmented and filtered to focus on the corn foliage using Facebook's Segment Anything Model 2. These images show an example of the original image, image mask and filtered image transition used to augment the training images.


Model Implementation

CNN Architecture

The project is framed as a supervised regression task — predicting a continuous SLB severity score (0–9 in 0.5 increments) from a single drone-collected RGB image. Models are evaluated on MAE, R², and snap-score accuracy (prediction rounded to nearest valid label). Data was split 70% training / 15% validation / 15% test.

A baseline custom 2-layer CNN with Adam optimizer and MSE regression loss established a starting point (MAE 1.17, R² 0.18). Custom feature engineering — adding an edge map channel and a "brown-surrounded-by-green" pixel mask — improved the baseline to R² 0.245 by directing the network toward disease-relevant regions.

Using state of art pre-built CNNs produced a major performance jump. EfficientNet V2 Small and Large were trained on the dataset, and a pre-trained Vision Transformer (ViT) was evaluated both as-is and with a novel directed attention mechanism — a custom loss term that penalizes attention weights misaligned with the brown-leaf disease mask, encouraging the model to focus on lesion regions. Grad-CAM visualizations confirmed the model was attending to disease-relevant areas.

Best model results (EfficientNet V2 Small):

Model comparison:


Tech Stack