L-ACDD: A Lightweight Automatic Crop Disease Detection System

Agriculture is a cornerstone of human civilization, but crop-related diseases adversely influence agricultural productivity, impacting farmers worldwide. Timely and effective disease detection and spread estimation are crucial for optimal plantation management. In recent years, deep learning-based classifiers have emerged as highly efficient for this purpose. However, deep learning methods have high computational complexity when used for classification purposes. Hence the proposed work exploits the capabilities of the Deep Convolutional Neural Network (DCNN) named Residual Network50 by modifying it to perform as a feature extractor instead of a classifier. The work presents a DCNN-based Lightweight Automatic Crop Diseases Detection (L-ACDD) System. We have employed a ResNet50-based Feature Extractor (ResNetFExt) to extract useful features from the given dataset. We have combined the extracted feature set with machine learning classifiers: Random Forest (RF), Support Vector Machine (SVM), and Logistic Regression (LR), individually, at the backend. The three developed systems have been trained and tested using two publicly available Plant Village Datasets (PVD), comprising 14 classes (PVD_14) and 38 classes (PVD_38), resulting in a total of six individual implemented systems. The obtained results clearly reveal that, among all these implemented systems, ResNet50 coupled with SVM demonstrates superior performance, with accuracies of 97.9% for the PVD_14 and 99.1% for the PVD_38.