1. Introduction
Food shortages may increase in many regions of the world. Coupled with pests and crop failures, food prices have soared. A lot of people may face severe hunger and death. In order to solve the food shortage, it is necessary to ensure the food security and sustainability. Due to pests, diseases [
1,
2], and lack of horticultural expertise [
3,
4,
5], food yield loss is greater than 50% [
6]. Food security is increasingly affected by crop production [
7]. With the increase of agricultural intensification and the continuous strengthening of the agricultural industry chain, the risks related to viruses and pollution will increase. For the goal of global food security and sustainable development, by 2050, the current demand of crop disease detection needs to increase by 50% [
8].
The traditional method of plant disease detection is manual inspection by farmers or experts. The method of plant disease diagnosis through optical observation of the symptoms on plant leaves incorporates a significantly high degree of complexity [
2]. The method laboratory-based such as polymerase chain reaction (PCR), immunofluorescence (IF), and fluorescence in-situhybridization (FISH) require professional laboratory equipment and mass sampling work [
9]. Due to this complexity and to the large number of cultivated plants and their existing phytopathological problems, manual plant disease detection can be time-consuming and expensive [
10]. By contrast, images under analysis were obtained by employing cameras operating in the visible portion of the electromagnetic spectrum (400–700 nm). In this way, costly equipment or trained personnel are not required for obtaining the input data [
11]. Therefore, future users of the developed protocol can acquire data through affordable/cost-effective, portable (thus in situ), and rapid means. With the development of computational systems in recent years, and in particular Graphical Processing Units (GPU) embedded processors, Convolutional Neural Networks (CNNs) [
12] is often applied for image classification.
CNNs belong to a stackable feedforward neural network community [
12]. The method of image classification through multi-layer CNNs is also called deep learning [
13,
14,
15]. CNNs have good characterization learning ability, so they are mostly used for feature extraction, and the extracted features have the characteristics of translation invariance. The research on CNNs began in the 1980s and 1990s, and the time delay network and LeNet-5 were the earliest CNNs [
12]. For a convolution operation, the essence is a traversal of the convolution kernel on the feature image. The convolution kernel will multiply and add the value at the corresponding position of the input feature image. In recent years, CNNs have been increasingly incorporated in plant phenotyping concepts. They have been very successful in modeling complicated systems, owing to their ability of distinguishing patterns and extracting regularities from data. Examples further extend to the variety identification in seeds [
16] and in intact plants by using leaves [
17]. Some research [
18] collected and published the datasets of plant diseases that provided data sources for other methods. There are also some works [
10] using image segmentation technology to separate the foreground and the background that can further improve the classification accuracy, and also solve the problem of poor performance on the online test.
Although the above deep learning methods have good performance related to plant disease detection, they still have the following problems. The first is the common problem of deep learning; model training requires a large amount of manually labeled datasets. The above-mentioned methods are currently based on the support of a large amount of data. Each category requires more than 1000 pictures. Data collection and marking require manpower and time. There is not enough data to support network training for plant pathogen variables in time, space, and genotype [
19]. The second problem is that many of the more than 700 known plant viruses cause devastating diseases and often have wide host ranges. Barley yellow dwarf viruses (BYDV), for example, are distributed worldwide and infect over 150 species of the Poaceae, including most of the staple cereals—wheat, barley, oats, rye, rice, and maize [
19]. It is unrealistic to identify all plant diseases at once through one task. However, the emergence of new tasks requires retraining the network. The above methods all limit the total number of categories for specific classification of several plant diseases. Every time a new task is encountered, based on traditional deep learning methods, it is necessary to rearrange the data and train the network to adapt to the task. For different sample numbers and image sizes, professional knowledge is needed to fine-tune the hyperparameters in the network structure. The last problem is the poor interpretability of the method. Compared with manual detection, experts can provide the basis for plant disease detection such as Oval-shaped irregular brown spots appearing on the leaves of plants with rust, and the leaf color on the leaves gradually becoming lighter, and, using fluorescence imaging, temporal and spatial variations of chlorophyll fluorescence were analyzed for precise detection of leaf rust and powdery mildew infections in wheat leaves at 470 nm [
9]. Although the deep-learning-based methods show their effectiveness, it cannot explain their decisions and actions to human users. Therefore, the methods should give visual explanations to illustrate that our approach focuses on diseases’ classification. This paper proposed a meta-learning [
20] method to solve the challenge of plant disease detection.
Meta-learning is the science of systematically observing how different machine learning approaches perform on a wide range of learning tasks, and then learning from this experience, or meta-data, to learn new tasks much faster than otherwise possible [
21]. There is a lot of research on meta-learning. Meta-learning is transfer learning in a broad sense [
22], which chooses data from different sources to train the network so that the model has a good classification effect on all kinds of tasks. Few-shot learning [
23], which is the problem of making predictions based on a limited number of samples, is an important application direction of meta-learning. The network is trained through other multi-source and sufficient datasets, so that it can deal with the task with few training samples. There is a lot of research on few-shot learning. For example, CNAPS [
24] and Simple-CNAPS [
25] use forward propagation instead of back propagation to solve the problem of overfitting, modular adaptation method [
26], and Meta Fine-Tuning [
27], which is also called Cross-Domain Few-Shot learning, can be trained to perform both tasks across domains. There is also some research on metric learning [
28] such as MatchingNet [
29] and ProtoNet [
23] to solve the problem of insufficient samples and the poor performance of the classifier. Although there have been many works on few-shot learning, most of the works are more theoretically focused, and do not focus on specific applications. Based on the previous work, this paper applied meta-learning to plant diseases detection.
4. Discussion
Pests and diseases seriously threaten crop yields, leading to food shortages, e.g., more than 800 million people do not have adequate food; 1.3 billion live on less than
$1 a day and at least 10% of global food production is lost to plant disease [
19]. In order to combat the losses, the emerging plant disease needs to be detected before it has a large-scale impact on crop growth. This paper intends to propose a method that can detect plant diseases with few samples. The results showed that LFM-CNAPS proposed has an average accuracy of 93.9% on detecting unseen plant disease with only 25 annotated examples. The method RESNET18 + FC based on deep learning [
10] only has an accuracy of 20.0%. More specifically, when classifying the five diseases of tomato: bacterial spot, early blight, healthy, late blight, and leaf mold, the performance of LFM-CNAPS is much better than the other two methods. LFM-CNAPS can give visual explanations through TAM, similar to optical observation of the symptoms on plant leaves. As show in
Figure 3, Alstonia Scholari’s wormhole is presented. However, due to complexity, even experienced agronomists often fail to successfully diagnose specific diseases [
2]. For example, it is difficult for people to distinguish the corn with northern leaf blight from the corn with gray leaf spot. LFM-CNAPS adaptively adjusts and extracts the potential differences between the two categories by task adaptive processes. This feature will be more abstract rather than simple geometric features. On a commercial scale, evidently, a capital investment is initially required for adopting the employed approach [
52]. Nevertheless, the wide-ranging large-scale commercial applications can provide high returns through considerable improvements in process enhancement and cost reduction.
Limitations of the study are due to a single task format. All tests are based on using 25 samples for plant disease detection. Recommendations for further research are to deal with a different scale of samples for better classification results.