Deep Learning for Plant Disease Identification from Disease Region Images

Abstract

This paper proposes a deep learning (DL) plant disease identification approach at leaf surface level using image data of pathologically segmented disease region or region of interest (ROI). The DL model is an exceptional technique used in automatic plant disease identification that employs a series of convolutions for feature representation of the visible disease region, mainly characterized as the combination of the chlorotic, necrotic, and blurred (fuzzy) lesions. The majority of current DL model approaches apply whole leaf image data for which studies have shown its consequential tendencies of leading to irrelevant feature representations of the ROI. The effects of which are redundant feature learning and low classification performance. Consequently, some state-of-the-art deep learning methods practice using the segmented ROI image data, which does not necessarily follow the pathological disease inference. This study proposes an extended ROI (EROI) algorithm using pathological inference of the disease symptom to generate the segmented image data for improved feature representation in DL models. The segmentation algorithm is developed using soft computing techniques of color thresholding that follows an individual symptom color feature that resulted in the incorporation of all lesions. The results from three different pre-trained DL models AlexNet, ResNet, and VGG were used to ascertain the efficacy of the approach. The advantage of the proposed method is using EROI image data based on pathological disease analogy to implement state-of-the-art DL models to identify plant diseases. This work finds application in decision support systems for the automation of plant disease identification and other resource management practices in the field of precision agriculture.