Abstract:
Potato is one of the most crucial food security crops to the world and its production in
Bangladesh is a great input to the sustenance of the country as well as being a vital food
source to the people. However, its commercial cultivation is facing the imminent danger
of a number of leaf diseases, including the early blight, late blight, viral diseases and
insect damage. The diseases mentioned severely kill the yield and quality and the
diagnosis should be made early and accurate. Conventional manual detection systems
are prone to flaws and show a slow response rate, hence there is a dire need to have an
automated system that is scalable. This paper introduces a deep learning framework
used to detect and label five different conditions of potato leaves, which includes Early
Blight, Late Blight, Virus, Insect, and Healthy. A custom Convolutional Neural Network
(CNN) was created, and it was trained on more than one thousand primary images. The
system has been compared to four CNNs comprising VGG19, ResNet50, MobileNetV2
and InceptionV3 based on standard performance measures like accuracy, precision,
recall and F1-score. This model had a classification accuracy of 98.95% which makes it
have a great potential in precision agriculture. Knowledge of uncertainties in terms of
accuracy was ensured along with the development of visual interpretation techniques
that increase the interpretability and trust on the decisions made by the model. This
study not only supports the power of deep learn
