Do we really need deep CNN for plant diseases identification?

Highlights

• The information extraction ability of shallow CNN was explored.

• The shallow CNN and classic machine learning classifiers were combined.

• The proposed methods outperform other deep models on several indicators.

Abstract

Timely identification of plant diseases plays crucial roles in the management and decision-making to protect the agricultural yield and quality. In this research field, there have been so many efforts focused on deep learning, namely deep CNN. The CNN is powerful and essential for image processing; however, do we really need deep CNN for plant diseases identification, cannot the shallow CNN extract enough information? We proposed two methods namely SCNN-KSVM (Shallow CNN with Kernel SVM) and SCNN-RF (Shallow CNN with Random Forest) to solve this confusion. The comparison experiments with other deep learning models were carried out on three different datasets. The results show that the SCNN-KSVM and SCNN-RF outperform other pretrained deep models on the indicators of precision, recall, and F1-score, with fewer parameters. The combination of shallow CNN and classic machine learning classification algorithm is a positive attempt to deal with the plant diseases identification in a simple manner.

Introduction

The plant diseases are important issues in agricultural production, if they cannot be identified in time, there will be a negative impact on the yield and quality of crops (Sethy et al., 2020). As known, the early detection and warning are basis of effective prevention and control for plant diseases, which play crucial roles in the management and decision-making (Fang and Ramasamy, 2015). However, the visual observations by experts or experienced farmers are still the primary approach for the plant diseases detection in many countries and areas. This traditional approach has many disadvantages, e.g. manual observations in large farm will be time-consuming and the high-frequency expert consultations will be too expensive. So, the automatic identification of plant diseases is of great realistic significance, which aims at detecting the symptoms of plant diseases as soon as they appear on leaves (Singh and Misra, 2017, Pooja et al., 2017).

Plenty of previous works have considered this problem by processing plant leaf images and designing some particular classifier to categorize the samples into healthy or diseased images. The reason to select plant leaf images as analysed data is that the plant leaves are generally the first area for symptoms of most plant diseases to appear (Barbedo, 2016, Barbedo, 2019). With the help of computer science and technology, there are two main categories of methods: classical machine learning and deep learning. The classical machine learning algorithms used for plant diseases identification include k-nearest neighbours (KNN) (Singh and Kaur, 2018), support vector machine (SVM) (Naik and Sivappagari, 2016), random forest (RF) (Chaudhary et al., 2016), etc. But these classical approaches despond heavily on the hand-designed features by various ways, such as HOG (histogram of oriented gradient), SIFT (scale-invariant feature transform), Gabor transform, PCA (principal components analysis), etc. On the other hand, the deep learning technology, particularly the convolutional neural network (CNN), have received excessive attentions to deal with the agricultural images, such as plant diseases and crop pests (Kamilaris and Prenafeta-Boldú, 2018, Li and Chao, 2020, Chen et al., 2020, Li and Yang, 2020, Thenmozhi and Reddy, 2019).

The deep CNN models have dominated in numerous fields of image processing research (Too et al., 2019). This phenomenon should be attributed to the powerful image processing ability of CNN, which can automatically extract features by the stack of convolutional layers. As known, the shallow convolutional layers in CNN can extract the primary features, e.g. edges, textures, colors, etc. With the deepening of CNN layers, the extracted features will become more and more abstract. Hence, we want to explore whether the deep CNN is necessary for plant diseases identification, cannot the shallow CNN provide enough features to deal with?

The main motivation of this study is to consider the future practical applications with intelligent algorithms. At present, many researches or algorithms are implemented on the high-powered hardware, such as GPUs and servers. To obtain better performance, the depth of the network becomes deeper with amounts of computational parameters, which consumes a lot of hardware resources and running time. But the high-powered hardware may not be suitable for the field applications, lacking of the convenience. In contrast, the embedded applications are more promising, due to the convenience and low cost (de Castro et al., 2020). Besides, the applications on smart phone running intelligent classification algorithm are also a promising trend, considering convenience, low power consumption, and low computation cost (Tao et al., 2020). In view of the above considerations, we want to explore the ability of shallow CNN to deal with the plant diseases identification.

In this paper, we proposed two methods namely SCNN-KSVM (Shallow CNN with Kernel SVM) and SCNN-RF (Shallow CNN with Random Forest) to deal with the plant diseases identification. The features of plant images are automatically extracted from the shallow CNN and then fed to the classical machine learning algorithms, such as Kernel SVM and Random Forest. By comparing the machine learning classification algorithms, the SVM algorithm is an important classification algorithm with ability to avoid the problem of overfitting, which is also one of the most frequently used machine learning classification algorithms (Liakos et al., 2018). Simultaneously, the Random Forest algorithm has excellent generalization ability and fast training speed (Liang et al., 2020). Thus, the Kernel SVM and Random Forest are selected to classify the extracted feature by the shallow CNN, which was excerpted from the pretrained VGG-16 model without any new trainable parameters. Our methods were compared with other deep learning models on three different datasets. It turned out that our method outperforms other deep learning models on the indicators of precision, recall, and F1-score, with fewer parameters.