Potato Leaf Disease Detection Using MultiNet: A Deep Neural Network with
Multi-Scale Feature Fusion
Abstract
Potatoes play a significant role in Bangladesh’s agriculture, ranking
third after rice and wheat, and contributing substantially to the
nation’s economy. However, leaf diseases pose significant challenges to
potato growth and quality despite Bangladesh being a major global potato
producer. Detecting diseases in a timely manner is essential for
enhancing productivity and advancing agricultural digitization. This
study aims to utilize machine-learning algorithms on a limited dataset
of potato leaf images for disease detection. The research focuses on
classifying potato leaves into distinct disease groups using The Plant
Village Dataset. A novel framework called “MultiNet” is introduced,
which employs transfer learning to effectively categorize various potato
leaf diseases. “MultiNet” facilitates rapid and accurate diagnostics
through two multi-class classifications: (1) Early Blight, Late Blight,
and Healthy (3 classes); and (2) Early Blight Normal, Early Blight
Serious, Late Blight, Healthy, Insect infected, LeafRoll Virus infected,
and Virus infected (7 classes). The system integrates three pre-trained
models—ResNet50, DenseNet-201, and VGG16—to extract features from
images, followed by concatenation to create a hybrid structure. The
results demonstrate an impressive overall classification accuracy of
99.83% for three classes and 98% for seven classes. These findings
underscore the potential impact and usefulness of the “MultiNet”
framework in agriculture, suggesting promising applications within the
sector.