Severity Classification of Philippine Downy Mildew in Corn using Infrared Thermography and Deep Learning

Philippine Downy Mildew is one of the most damaging diseases affecting maize (corn) globally, causing significant damage to crops, lower yields, and increased production costs. This study proposes a plant disease detection and severity classification system that uses infrared thermography and deep learning. A novel dataset of 2279 corn thermograms was collected in the Institute of Plant Breeding field environment in Los Baños, Laguna, Philippines. Philippine Downy Mildew is categorized based on infection severity: Highly Resistant (0%-5%), Moderately Resistant (10%-20%), Moderately Susceptible (30%-50%), and Susceptible (60%- 100%). 3 CNN (Convolutional Neural Network) architectural models were compared for their performance with thermogram input. The models developed through customizing CNN architectures using ensemble learning include Model A (two-stage binary classification with NASNetMobile and MobileNetV2), Model B (multi-classification with EfficientNetB1 and NASNetMobile), and Model C (integrated Model B with binary classification using MobileNetV2). During the training of the customized CNN models, the use of transfer learning aided the increase in multiclassification accuracy. NASNetMobile obtained an accuracy of 93.8% for binary and 90.7% for multiclass, a two-class binary classification of MobileNetV2 obtained 95.8% and 96.9% respectively, and a multiclass EfficientNetB1 obtained 91.1%. During the actual field testing, models A, B, and C achieved an accuracy of 56.92%, 63.08%, and 76.92% respectively. This study demonstrates the potential of implementing infrared thermography with deep learning algorithms in agriculture for non-invasive plant disease severity rating.