Automatic Detection and Severity Assessment of Pepper Bacterial Spot Disease via MultiModels Based on Convolutional Neural Networks

Automatic Detection and Severity Assessment of Pepper Bacterial Spot Disease via MultiModels Based on Convolutional Neural Networks

Qiufeng Wu (Northeast Agricultural University, China), Miaomiao Ji (Northeast Agricultural University, China) and Zhao Deng (Northeast Agricultual University, China)
DOI: 10.4018/IJAEIS.2020040103

Abstract

Pepper bacterial spot disease caused by Xanthomonas campestris is the most common pepper bacterial disease, which ultimately reduces productivity and quality of products. This work uses deep convolutional neural networks (CNNs) to serve fine-grained pepper bacterial spot disease severity classification tasks. The pepper bacterial spot disease leaf images collected from the PlantVillage dataset are further annotated by botanists and split into healthy samples (label1), general samples (label2), and serious samples (label3). To extract more effective and discriminative features, an integrated neural network denoted as MultiModel_VGR is proposed for automatic detection and severity assessment of pepper bacterial spot disease, which is based on three powerful and popular deep learning architectures, namely VGGNet, GoogLeNet and ResNet. Compared with state-of-the-art single CNN architectures and binary-integrated MultiModels, MultiModel_VGR yields the best overall accuracy of 95.34% on the hold-out test dataset, which may have great potential in crop disease control for modern agriculture.
Article Preview
Top

1. Introduction

Pepper, as one of the most widely cultivated economical crop throughout the world, plays an important role in the world's agricultural production and trade. For decades, pepper has been severely affected by various diseases, such as red anthracnose, bacterial spot and black mold, etc. Among these diseases, pepper bacterial spot disease caused by Xanthomonas campestris is the most common pepper bacterial disease, which can ultimately reduce productivity and quantitative losses in crop yield. A very small number of diseased leaves tend to spread the infection to the whole batch of crops and cause epidemic all over the field, which is undoubtedly devastating. Timely information on the harm degree of crop disease can guide agricultural producers to take necessary and more targeted measures to avoid additional financial losses. Therefore, it is urgent to develop an early and accurate method for pepper bacterial spot disease detection and severity assessment.

The existing method for crop diseases detection and severity assessment is simply naked eye observation by farmers in the field with the guidance of plant pathologists, whose process of diagnosis is not only subjective but also time-consuming and laborious. Advances in science and technology now make it possible for computer vision approaches to assist us in automatic detection and severity assessment of crop diseases tasks. CNNs can locate important features itself and automatically learn appropriate features from training dataset instead of manual feature extraction. They have such excellent generalization ability and robustness that they excel in diverse areas, such as signal processing (Xie et al., 2017), road crack detection (Zhang et al., 2016) and biomedical image analysis (Zhou et al., 2017). CNNs also have been successfully applied in the field of agriculture, including diagnosis of crop diseases (Zhang et al., 2019a, 2019b), recognition of weeds (Philipp et al., 2018), selection of fine seeds (Wang & Cheng, 2016), pest identification (Cheng et al., 2017), fruit counting (Rahnemoonfar & Sheppard, 2017) and land cover classification (Kussul et al., 2017), etc. Although remarkable performances have been achieved in normal crop diseases classification, it is still hard to distinguish diseases with subtle discrimination. Compared with classification among different crop diseases, pepper bacterial spot disease severity extent classification is much more challenging, as there exist large intra-class similarity and small inter-class variance in pathological symptoms. As shown in Figure 1, in the early stage of infection, Xanthomonas campestris bacterium attack the pepper leaf causing one or several light-colored spots. The serious indication of the disease on leaflets consists of a large number of brownish lesions which are not delimited in size and enlarge rapidly under favourable conditions.

Fine-grained image classification refers as discriminating the sub-categories sharing one common basic-level category through digital images (Sumbul et al., 2017). Pepper bacterial spot disease detection and severity assessment belongs to the category of fine-grained classification task and our schedule is aimed to directly boost up the accuracy through more effective and discriminative features and representations extracted by an effective algorithm. Inspired by the great success of CNNs-based methods in image classification as well as other applications, the integrated model in our work, denoted as MultiModel_VGR, is just based on the three powerful and popular deep learning architectures, namely VGGNet (Simonyan & Zisserman, 2014), GoogLeNet (Szegedy et al., 2014) and ResNet (He et al., 2015), where the input data is propagated forward to each basic single model and then input to the classifier after concatenation. In addition, it is easier to fuse the features extracted from the convolutional layers of each single model by way of transfer learning.

Figure 1.

Illustration of the pepper bacterial spot disease: healthy samples, general samples and serious samples

IJAEIS.2020040103.f01

Complete Article List

Search this Journal:
Reset
Open Access Articles
Volume 12: 4 Issues (2021): Forthcoming, Available for Pre-Order
Volume 11: 4 Issues (2020)
Volume 10: 4 Issues (2019)
Volume 9: 4 Issues (2018)
Volume 8: 4 Issues (2017)
Volume 7: 4 Issues (2016)
Volume 6: 4 Issues (2015)
Volume 5: 4 Issues (2014)
Volume 4: 4 Issues (2013)
Volume 3: 2 Issues (2012)
Volume 2: 2 Issues (2011)
Volume 1: 2 Issues (2010)
View Complete Journal Contents Listing