Lightweight and Explainable Deep Learning Models for Pomegranate Disease Diagnosis: A Case Study with Maharashtra Observations

Abstract

Pomegranate is a major fruit crop in Maharashtra with major cultivation marked in the districts Solapur, Satara, Sangali, Ahmednagar and Latur. But farmers in these districts face significant challenges from disease outbreaks. Traditional methods of disease diagnosis are often time-consuming and inefficient in generating accurate results. In the recent year, deep learning has emerged as an effective solution for automation in disease detection. However, deep learning models are often computationally heavy and lack interpretability, restricting their real-world applications. But the lite versions of these models are exceptionally small in size and hence perform well in real time. So, this study tests the effectiveness of MobileNetV3-Lite and EfficientNet-Lite models on the available dataset comprising of 5,099 images of pomegranate fruit from various regions of India; this study focuses on diseases prevalent in Maharashtra orchards to ensure contextual relevance. Further, integration of Grad-CAM visualizes affected regions within image by creating heatmaps and hence providing insight into the reasoning behind the model’s predictions. The results show that EfficientNet achieved highest classification accuracy of 99.50%, in contrast MobileNetV3 achieved marginally lower accuracy 98.48% but its lite model exhibited a significant smaller size 6.17 MB along with inference time of 0.019 s/image, making it more suitable for real-time and low resource agricultural applications.