Abstract
Smart agriculture is the application of modern information and communication technologies (ICT) to agriculture, leading to what we might call a third green revolution. These include object detection and classification such as plants, leaves, weeds, fruits as well as animals and pests in the agricultural domain. Object detection, one of the most fundamental and difficult issues in computer vision has attracted a lot of attention lately. Its evolution over the previous two decades can be seen as the pinnacle of computer vision advancement. The detection of objects can be done via digital image processing. Machine learning has achieved significant advances in the field of digital image processing in current years, significantly outperforming previous techniques. One of the techniques that is popular is Few-Shot Learning (FSL). FSL is a type of meta-learning in which a learner is given practice on several related tasks during the meta-training phase to be able to generalize successfully to new but related activities with a limited number of instances during the meta-testing phase. Here, the application of FSL in smart agriculture, with particular in the detection and classification is reported. The aim is to review the state of the art of currently available FSL models, networks, classifications, and offer some insights into possible future avenues of research. It is found that FSL shows a higher accuracy of 99.48% in vegetable disease recognition on a limited dataset. It is also shown that FSL is reliable to use with very few instances and less training time.
Few-Short Learning in agriculture
In the realm of computer vision, detecting small objects such as plants, leaves, weeds, fruits, and pests is a critical study topic. It is a technology that employs computer vision equipment to acquire photos to determine whether the obtained images include weeds, pests or even plant diseases (Lee et al., 2017). Computer vision-based object detection technology is currently being used in agriculture and has partially replaced traditional naked-eye identification (Liu and Wang, 2021). During the growth process, pests, and weeds frequently harm crops which leads to various diseases. They will become more dangerous if control is not done in a timely or effective manner. This will have an impact on crop yields and result in significant losses in agricultural income. People rely on their experiences while using traditional detection methods (Zhang S. et al., 2021).
Numerous agricultural plants, weeds, fruits, pests, and diseases have similar morphology, making it difficult to recognize them. Only judging by human eye observation is time-consuming, debilitating, biased, imprecise, and costly (Ashtiani et al., 2021). Nonprofessional agricultural employees will be harmed by their personal experiences since they will be unable to appropriately identify the species. As a result, pest control is an important aspect of agricultural productivity, and detecting agricultural pests is critical for agricultural development (Zhang S. et al., 2021).
As a result, an accurate detection technique is very desirable, and FSL is an approach that is precise in object detection. FSL learning is a sub-area in machine learning. FSL is a sort of machine learning problem (defined by E, T, and P), in which E, Experience, comprises a small number of supervised instances for T, Task, measured by P, Performance (Wang et al., 2020). FSL is used to quickly train the model on a little amount of tagged data (Fei-Fei et al., 2006). The fundamental concept behind FSL is comparable to how humans learn: it makes use of previously learned knowledge to pick up new skills while minimizing the quantity of training data needed. Research findings (Duan et al., 2021) demonstrate that FSL can address several machine learning's drawbacks, such as the need for huge datasets to be collected and labeled, by dramatically lowering the quantity of data necessary to train machine learning models. Due to the small number of datasets, the data are divided into two splits [train and test (without validation) or train and validation (without test)] in many types of research in the field of artificial intelligence, which can be a weakness. The data set is divided into two subsets because, in small data sets, an additional split could result in a smaller training set that is more susceptible to overfitting (Ashtiani et al., 2021).
A framework put forth by Fink (2004) allows object classifiers to be learned from a single example, which drastically reduces the amount of data gathering required for image classification tasks. Additionally, FSL enables the training of models that are appropriate for a few uncommon circumstances (Duan et al., 2021). A good example is the significant advancements in drug development that have been made as a result of the most recent FSL research advancements (Jiménez-Luna et al., 2020). Clinical biological data for a given drug are extremely scarce when predicting whether it is harmful, which frequently makes use of classic machine learning algorithms difficult.
FSL is widely used for scene classification (Alajaji et al., 2020), text classification (Muthukumar, 2021), image classification (Li X. et al., 2020) and image retrieval (Zhong Q. et al., 2020). Even though software and computational technologies have long been utilized for detection in agriculture, FSL's capacity to identify them with limited training data makes them suitable for the task. This is ground-breaking because, they can be detected under various conditions of lighting, orientation, and background (Nuthalapati and Tunga, 2021).
FSL has been proven in agriculture to be an effective method for plant disease identification (Argüeso et al., 2020). Recent FSL approaches in agriculture are mainly used for plant disease detection (Wang and Wang, 2019; Li and Chao, 2021), fruit detection and classification (Janarthan et al., 2020; Ng et al., 2022), leaf identification and classification (Afifi et al., 2020; Jadon, 2020; Tassis and Krohling, 2022) and pest detection (Li and Yang, 2020, 2021; Nuthalapati and Tunga, 2021) using the meta-train set which is subsequently used to build embeddings for the samples in the meta-test set.
This article reviews the limitations of deep learning, advantages of FSL, types of FSL networks, their applications in the agricultural sector and their corresponding findings.
Limitations of deep learning
In agriculture, deep learning-based automatic classification of several categories is a hot topic of research (Too et al., 2019; Li Y. et al., 2020). Yet, in general, people still lean on non-automated classification by experts for the classification of distinct species. This is due in part to the fact that deep learning networks based on Convolutional Neural Networks (CNNs) demand thousands of labeled examples per target category for training, and labeling samples on such a large scale necessitates domain experts (Krizhevsky et al., 2017).
Furthermore, following training, the number of categories that a trained CNN-based model can recognize remains constant. To increase the number of categories that the network can recognize, it must be fine-tuned by adding fresh examples from other classes (Chen et al., 2019). Fine-tuning is a transfer learning approach which is used to benefit from the pre-trained network by adjusting the parameters to the data set. Due to the pre-established weights in a pre-trained network, fine-tuning is quicker than training from start (Ashtiani et al., 2021). Also, adequate data is required during training to avoid the network from overfitting (Gidaris and Komodakis, 2018) whereas humans can acquire new skills with little to no guidance using one or a few examples only (Lake et al., 2011).
Advantages of Few-Shot Learning
Humans can generalize new information based on a small number of examples, whereas artificial intelligence typically needs thousands of examples to produce comparable results (Duan et al., 2021). Researchers sought to create a machine learning model that, after learning a significant amount of data for some categories, could quickly learn a new category with only a few sample data, drawing inspiration from humans' rapid learning abilities. This is the issue that FSL seeks to address. This task only offers a small number of observable examples, much like the production of character samples (Lake et al., 2015).
FSL has lately received a lot of attention as a way for networks to learn from a few samples. FSL tries to solve the challenge of classification with a small number of training examples, and it is gaining traction in a variety of domains (Lake et al., 2011). In FSL, there are two sets of labeled image data which are meta-train and meta-test, with image classes that are mutually exclusive in both sets (Argüeso et al., 2020). The goal is to develop a classifier on the visual classes in the meta-test set that can classify a particular query sample even with very few labeled examples using the data in the meta-train set and learn transferrable information (Altae-Tran et al., 2017).
Additionally, cross-domain FSL is defined as when the domain of visual classes in the meta-train set differs from the domain of visual classes in the meta-test set (Argüeso et al., 2020). One of the primary issues in cross-domain FSL is feature distribution differences between domains. Likewise, both the meta-train and meta-test sets in mixed-domain FSL contain classes from many domains, whereas single-domain FSL (Li and Yang, 2021) only contains instances from a single domain.
According to research (Duan et al., 2021), FSL can learn about uncommon circumstances where it is challenging to find information about supervision by combining one or a few examples with previously learned knowledge. This can solve the issue of the machine learning algorithm's need for thousands of supervised examples to guarantee the model's generalizability. Additionally, FSL can help expedite the training cycle and lower the high cost of data preparation (Parnami and Lee, 2022).
Types of networks in Few-Shot Learning
According to the traditional paradigm, an algorithm learns whether task performance gets better with practice when given a specific task. The Meta-Learning paradigm consists of several tasks. An algorithm that is learning to determine whether its efficiency at each task increases with practice and the number of tasks is called a meta-learning algorithm (Chen et al., 2021). Numerous Meta-Learning techniques have been released by researchers in recent years to address FSL classification issues. Metric-Learning (Yang and Jin, 2006) and Gradient-Based Meta-Learning (Khodak et al., 2019) methods can be used to categorize them all into two main classes.
Metric-learning is the term used to describe the process of learning a distance function over objects (Chen et al., 2021). Algorithms using metrics learn to compare data samples in general. They categorize query samples for an FSL classification problem according to how similar they are to the support samples. A base-learner and a meta-learner are created for the gradient-based technique. A base-learner is a model that is initialized and taught within each episode by the meta-learner, whereas a meta-learner is a model that learns across episodes (Cao et al., 2019). Table 1 shows the meta-learning algorithms with its category respectively.
Table 1
| Networks | Category | Embedding function for testing | Embedding function for training |
|---|---|---|---|
| Model-agnostic meta learning (Wang and Wang, 2019) | Gradient-based meta learning | Deep neural network | Deep neural network |
| Matching network (Wang and Wang, 2019) | Metric learning | CNN, LSTM | CNN, biLSTM |
| Prototypical network (Yavari, 2020) | Metric learning | CNN | CNN |
| Relation network (Yavari, 2020) | Metric learning | CNN | CNN |
| Siamese network (Yavari, 2020) | Metric learning | CNN | CNN |
Types of meta learning algorithms.
One of the most well-liked meta-learning algorithms, Model-Agnostic Meta-Learning (MAML) has made significant advancements in the field of meta-learning research. Furthermore, meta-learning is a two-part learning process where the model first prepares to learn new tasks and then actually learns the new task from a few samples in the second stage (Yavari, 2020). The general goal of MAML is to identify suitable starting weight values so that the model can quickly pick up new tasks.
The first Metric-Learning method created to address FSL issues was Matching Networks (Chen et al., 2021). When using the Matching Networks approach to resolve an FSL job, a large base dataset is required (Li X. et al., 2020). This dataset is divided into episodes. Every image from the query and support sets is given to a CNN, which produces embeddings for each one. The SoftMax of the cosine distance between each query image's embeddings and the support-set embeddings is used to classify each image. It also uses Long Short-Term Memory (LSTM) for testing and Bidirectional Long Short-Term Memory (biLSTM) for training (Ye et al., 2020). The resulting classification's Cross-Entropy Loss is backpropagated through the CNN. Matching Networks can learn to construct picture embeddings in this way. Matching Networks can categorize photographs using this method without having any special prior knowledge of classes (Vinyals et al., 2016). Simply comparing several instances of the classes is used for everything.
Prototypical Networks learn a metric space in which each class may be classified by computing distances between prototype representations (Yavari, 2020). They represent a more straightforward inductive bias, which is advantageous in this low-data environment. The notion behind Prototypical Networks is that each class has an embedding in which points cluster around a single prototype representation (Pan et al., 2019). To accomplish so, a neural network is used to learn a non-linear mapping of the input into an embedding space, and a class's prototype is taken to be the mean of its embedding space support set. The next step is to classify an embedded query point by simply identifying the closest class prototype (Ji et al., 2020).
Another straightforward and efficient technique for both one-shot and FSL that adheres to the metric-learning paradigm is the Relation Network (Yavari, 2020). Relation Network uses an embedding function f to extract the features from the inputs, just like the Prototypical and Siamese Networks (Sung et al., 2018). As an embedding function, CNN is utilized because the inputs are images. The Relation Network includes a relation function in addition to the embedding function. This relation function evaluates how closely the query sample is connected to the various classes in the support set (Sun et al., 2018).
A Siamese Network is made up of two networks that receive different inputs but are connected at the top by an energy function (Yavari, 2020). This function calculates a metric between each side's highest-level feature representation. The dimensions of the twin networks are inextricably linked (Dong and Shen, 2018). The network's structure is duplicated in both the top and bottom parts, resulting in twin networks with shared weight matrices at each layer. The Siamese Network's goal is to evaluate two photos and determine whether they are similar (Melekhov et al., 2016). The process for comparing a Siamese network is quite straightforward.
Other methods for small data training
There are also methods other than FSL for small data training. Knowledge may be successfully exploited in sophisticated applications by being appropriately organized and represented in knowledge graphs (Chen et al., 2020). Knowledge graphs can efficiently manage, mine, and organize knowledge from massive amounts of data, enhancing the quality of information services and delivering more intelligent services to users. It is used to store connected descriptions and details about things like events, objects, actual circumstances, and theoretical or abstract ideas (Pujara et al., 2013). This approach simultaneously encodes semantics underlying the data collection in addition to serving as data storage. The arrangement and structuring of significant data points from the data set to integrate information gathered from multiple sources is the main purpose of the knowledge graphs model. Labels are added to a knowledge graph to associate certain meanings. Nodes and edges are the two basic parts of a graph (Wang et al., 2017). Nodes are made up of two or more objects, and edges show how they are related to one another.
By using enough prior knowledge from related fields to complete new tasks in the target domain, transfer learning simulates the human visual system despite the size of the data set (Shao et al., 2014). In transfer learning, the source domain and the target domain are two different sorts of domains that can both benefit from the training and testing data (Weiss et al., 2016). The source domain contains training examples, which are under a different distribution from the target domain data, and the target domain contains the testing instances, which are the classification system's task. For a transfer learning task, there is typically just one target domain, although there may be one or more source domains (Torrey and Shavlik, 2010).
Another popular method is self-supervised learning (SSL). SSL has becoming more common since it can save on the expense of annotating huge datasets (Jaiswal et al., 2020). It has the ability to employ the learnt representations for numerous downstream tasks and use self-defined pseudo labels as supervision. Particularly, contrastive learning has lately taken center stage in self-supervised learning for computer vision, natural language processing (NLP), and other fields. It attempts to push embeddings from various samples away while trying to embed augmented copies of the same sample closely together (Zhai et al., 2019).
Methodology
A literature search has been carried out by querying Google Scholar, IEEE Xplore, ScienceDirect, SpringerLink, ACM Digital Library and Scopus. According to studies, Google Scholar's reach and coverage are greater than those of similar academic websites. Then, using exclusion criteria, a small number of databases is chosen to search for pertinent papers. The search was conducted using the terms “few shot learning”, “smart agriculture”, “object detection”, “few shot classification”, “one shot learning”, “low shot learning”, “Meta-Learning”, “Prototypical Network”, “Siamese Network”, “MAML”, “Relation Network”, and “Matching Network”.
The search was also limited to recent research papers in the years 2010 and above. Repeated things between search engine results were searched by employing a straightforward script based on the edit distance between article titles. Around 150 articles were returned by all the search engines with 58 articles in IEEE Xplore, 50 articles in ScienceDirect, 21 articles in Google Scholar, 15 articles in SpringerLink, and six articles in Scopus. The resulting papers were very diverse, and most of them fell outside the purview of the review. As a result, studies that had been published and that dealt with FSL in agriculture and reported noteworthy findings considering recent literature have been chosen. The search was restricted to keywords, titles, and abstracts.
Exclusion criteria were used to screen out works that had nothing to do with the literature review. The following standards are disregarded in this analysis: (a) The ensuing articles have nothing to do with FSL in agriculture, (b) Articles that are not written in English, (c) Articles that are duplicated in the databases, (d) Articles that do not have significant results and problems. This filtered to 30 articles and during the process, references of the selected papers were also looked to find other works that fit the inclusion criteria. Finally, the resulting articles were organized by considering the models and types of objects in the agriculture field.
Results and discussion
FSL was first used in the field of computer vision, where it produced promising results in picture classification. It does not require large samples during training, which reduces the time and cost of obtaining samples. As a result, it is now being welcomed in many fields such as smart agriculture (Yang et al., 2022). FSL has shown compelling results in smart agriculture such as plant disease classification, plant leaf classification and plant counting. Early diagnosis of plant diseases is crucial for preventing plagues and mitigating their consequences on crops.
Deep learning is used to create the most accurate automatic algorithms for identifying plant diseases using plant field pictures. Traditional machine learning and deep learning techniques, on the other hand, have certain drawbacks, one of which is the requirement for large-scale datasets for models to generalize well (Wang et al., 2020). These methods necessitate the collecting and annotation of huge image datasets, which is usually impossible to achieve due to technological or financial constraints. As a result, new strategies were presented with the goal of developing models that generalized well with fewer data. As a result, meta-learning approaches (Hospedales et al., 2021) and FSL methods, were proposed. The Table 2 shows the results of the findings.
Table 2
| References | Data source | Techniques | Accuracy | Purpose | Limitation |
|---|---|---|---|---|---|
| Wang and Wang (2019) | Flavia Dataset; Swedish Dataset; Leafsnap Dataset | Siamese Network; Simultaneous Two-way Convolutional Neural Network; Spatial Structure Optimizer (SSO); K-Nearest Neighbor (KNN) | Flavia = 95.32%; Swedish = 91.37%; Leafsnap = 91.75% | To handle a leaf classification problem with a small sample size using Siamese Network | Has to improve the generalization ability of the model |
| Karami et al. (2020) | Maize Plant Dataset | Modified CenterNet Architecture | >95% | To identify and count maize plant by using a smaller number of images in a field using FSL | Adapting generative adversarial network (GAN) in future |
| Ng et al. (2022) | Fruit Dataset | Meta-learning paradigm; Few-shot classification | >75% | To classify new varieties of fruits with only a few training examples | Not stated |
| Argüeso et al. (2020) | PlantVillage Dataset | CNN; SVM Classifier; Siamese Network; Few-shot and baseline transfer learning; Root Mean Square optimizer (RMSprop) | Triplet Loss; Accuracy = 91.4%; Precision = 93.8%; Recall = 92.6%; F-Score = 92.3%; Siamese; Accuracy = 90.6%; Precision = 93.8%; Recall = 93.1%; F-Score = 91.2% | To classify plant diseases to overcome unfeasibility of deep learning | Convergence problems |
| Zhong F. et al. (2020) | CUB Dataset; AWA2 Dataset; APY Dataset | Conditional Adversarial Autoencoders (CAAE); Kullback-Leibler (KL) Divergence Penalty | Zero Shot; Unseen = 55.4%; Seen = 51.6%; Harmonic = 53.4% | Recognition of Citrus aurantium L. diseases using few and zero shot learning | Only uses a discriminator to replace KL-divergence in variational autoencoders |
| Li and Chao (2021) | PlantVillage Dataset | Semi-supervised few-shot learning | 5-shot = 90%; 10-shot = 92.6% | To resolve plant leaf disease recognition problem using semi-supervised FSL | Did not consider wrong labels |
| Wang et al. (2021) | Self-Collected Dataset | Image text collaborative representation learning (ITC-Net) | Accuracy = 99.48%; Precision = 98.90%; Sensitivity = 98.78%; Specificity = 99.66% | Recognition of vegetable illness in complicated backgrounds | The feature spaces of image and text are still independent |
| Jadon (2020) | Mini-Leaves Diseases Dataset; Sugarcane Dataset | Stacked Siamese Matching (SSM) Net | Mini Leaves = 92.7%; Sugarcane = 94.3%; Comparison with VGG16; Mini Leaves = 0.91; Sugarcane = 0.90 | Mini-leaves diseases and sugarcane illness identification using combined methods of Siamese and Matching networks | Not stated |
| Hu et al. (2019) | Tea Leaf Disease Image Dataset | Conditional deep convolutional generative adversarial networks (C-DCGAN); Support vector machine (SVM) | 90% | Tea leaf disease identification with low shot learning | Low generalization performance |
| Janarthan et al. (2020) | Citrus Fruits and Leaves Dataset | Deep Siamese Network | 95.04% | To create a classification strategy for diagnosing citrus diseases | Not Stated |
| Afifi et al. (2020) | PlantVillage Dataset; Coffee Leaf Dataset | Triplet network; Deep adversarial metric learning (DAML) | 99% | Automatic identification of plant diseases with limited data | Focusing on the full image instead of focusing on the affected part |
| Tassis and Krohling (2022) | Leaf Dataset; Symptoms Dataset | Prototypical networks; Triplet networks; feature extraction | 1st Experiment = 96.03%; 2nd Experiment = 96.72%; 3rd Experiment = 93.25% | FSL in classification and severity assessment tasks with biotic stressors in coffee leaves | During training, there is a lack of performance evaluation of the approaches in unobserved classrooms |
| Li and Yang (2020) | National Bureau of Agricultural Insect Resources (NBAIR) Dataset; Li's dataset of natural scenes | Prototypical Network; CNN; Triplet loss; Terminal realization | 95.4% (National Bureau of Agricultural Insect Resources dataset); 96.2% (Li's dataset of natural scenes) | Few shot cotton pest recognition and terminal realization | Not stated |
| Nuthalapati and Tunga (2021) | Plant and Pests (PP) dataset; PlantVillage dataset; Plants in Wild (PiW) dataset; Cross dataset | Few-shot learning using transformer | Plant and Pests = 60.3%; PlantVillage = 88.5%; Plants in Wild = 93.3%; Cross = 86.1% | To classify diverse pests, plants, and illness | Not stated |
| Li and Yang (2021) | PlantVillage Dataset | Few-shot classification; CNN | Single domain,; Ntrain = Ntest = 3 = 90.4%,; Ntrain = Ntest = 5 = 82.5%.; Cross domain 1,; Ntrain = Ntest = 3 = 86.5%,; Ntrain = Ntest = 5 = 81%.; Cross domain 2,; Ntrain = Ntest = 3 is 54.8%,; Ntrain = Ntest = 5 is 43.9%. | Task-driven meta-learning few shot classification for pests and plants and to investigate single domain and cross-domain classification | Not stated |
| Zhang H. et al. (2021) | MiniImagenet Dataset; CUB-200-2011 Dataset; Flower102 Dataset | Absolute-relative Learning | 1-shot = 55.61%; 5-shot = 71.03% | Absolute-relative learning paradigm in supervised and unsupervised FSL | Only used semantic annotations as labels in training, and do not use them during testing |
| Liu et al. (2019) | ImageNet Dataset | Few-Shot Unsupervised Image-to-Image Translation (FUNIT); Generative Adversarial Networks (GAN) | Animal Face; 1-shot = 82.36%; 5-shot = 96.05%; North American Birds; 1-shot = 60.19%; 5-shot = 75.75% | Few-shot unsupervised image-to-image translation of animal faces, birds, flowers, and foods | FUNIT fails to achieve translation when the look of novel object classes differs considerably from that of the source classes |
Studies of few-shot learning in smart agriculture.
Research on identification has begun over two decades. Deep learning was used in these studies, and they were a huge success. However, there are a few challenges that deserve greater attention and research, including the demand for a big training set, training time, CNN architecture, pre-processing, feature extractions, and performance evaluation measures. Some of the experiments will not work in the wild due to camouflaging of the backgrounds (Liu et al., 2019; Kuzuhara et al., 2020; Yao et al., 2020).
To address these problems, authors have utilized machine learning in the identification of pests as this study makes the machine to think like humans. Machine learning has shown satisfactory results in most studies (Brunelli et al., 2019; Tachai et al., 2021; Zhang H. et al., 2021). However, machine learning also uses CNN feature extractions for training which shows improvement. Classifiers such as Support Vector Machine (SVM) is the best classifier that enhances the accuracy (Kasinathan et al., 2021).
To fill the gap between large training samples and time consumption, FSL was implemented. FSL has become popular in the computer vision field in both classification and detection. FSL was introduced very recently, and the results generated by this meta-learning approach are favorable. In the smart agriculture sector, FSL is widely used in plant disease identification (Wang and Wang, 2019; Li and Chao, 2021), plant counting (Karami et al., 2020), leaf classification (Afifi et al., 2020; Jadon, 2020; Tassis and Krohling, 2022) and fruit ripeness classification (Janarthan et al., 2020; Ng et al., 2022). The results for these experiments have exceeded 90%.
For plant leaf classification using the Siamese network, the accuracy of 95.32%, 91.37% and 91.75% were achieved for Flavia, Swedish and Leafsnap respectively (Wang and Wang, 2019). For maize plant counting, the overall precision was above 95% (Karami et al., 2020). For plant disease classification using triplet loss, the accuracy was 91.4% and using the Siamese network the accuracy was 90.6% (Argüeso et al., 2020). In vegetable disease recognition 99.48% accuracy was reached (Wang et al., 2021). For pest detection, have utilized a prototypical network and triplet loss with a terminal realization for cotton pests. They used two datasets for comparison and achieved 95.4 and 96.2% respectively (Li and Yang, 2020; Nuthalapati and Tunga, 2021).
However, some FSL techniques show lower accuracies due to their limitations. For fruit ripeness classification, they achieved 75% accuracy for only five samples (Ng et al., 2022). Another plant disease recognition has achieved an accuracy of around 55% for zero-shot (Zhong F. et al., 2020). For pests, one study achieved 60.3% (Nuthalapati and Tunga, 2021). For animal detection, the authors achieved 55.61 and 71.03% for 1 and 5 shots respectively due to no semantic annotations in testing (Zhang H. et al., 2021). Another bird detection study achieved 60.19 and 75.75% for 1 and 5 shots respectively (Liu et al., 2019).
Even though some research has lower accuracies, they are still satisfactory because FSL only uses very less instances in the training of 1, 2, 3, 5, and 10 images for detection and classification (Zhong F. et al., 2020). This proves that FSL is giving a good result with very less training samples compared to CNN which requires more training samples and training time.
Main findings and ways forward
Deep learning and machine learning is the process of utilizing algorithms to direct a computer in building a suitable model based on existing data and then using that model to evaluate novel situations. The models' applicability are significantly constrained by the massive amounts of labeled data that are often needed to train it because it has multiple parameters. Large amounts of tagged data are frequently exceedingly expensive, challenging, or even impossible to obtain. FSL is a crucial component of the development of machine learning since it is a crucial question if a good model can be created by training with only a minimal amount of labeled data. It may be applied to learn uncommon situations using a limited collection of characteristics, which is a method that can successfully address the issue of dataset imbalance in intrusion detection.
In recent years detection and classification in the agricultural domain has received a lot of attention. Beginning with deep learning and machine learning, they are showing excellent results in identifying the objects. This approach is very useful for limited datasets and time for detection as well as classification of various things including plant diseases and pests. Most of the authors used the current state of the art in FSL such as the Siamese network, Prototypical network, and Model-Agnostic Meta Learning network.
Most of the FSL approaches show favorable results of more than 90%. One research has reached 99.48% accuracy in vegetable disease detection using a self-collected dataset. Even though the accuracy is very excellent in this research, the authors stated that the feature spaces of image and text are still independent, and it needs to be improved in the future. Triplet network and deep adversarial metric learning have also given accuracies of 99% with full coffee leaf images. Besides FSL, one shot learning has also shown a very satisfying result of 95.32% using the Siamese network. The low generalization ability of the model is an issue in this research and the gradient descent in the training must be stopped at the appropriate point as well as adding a regularization term to the error function to smooth out the mappings.
Most of the results of plant datasets have higher accuracies than animal datasets. However, pest detection using a prototypical network has achieved 96.2% accuracy. Another pest detection only achieved 60.3% due to camouflaging images in the background which can be improved by focusing on the insects only. The least accuracies that have been obtained are with animal detection, 55.61% for one shot. However, when five shots were used, the accuracy increased to 71.03%. This proves that when more instances are used, the accuracy increases. Despite that, Few-Shot Unsupervised Image-to-Image Translation (FUNIT) has produced an accuracy of 82.36% for one shot and 96.05% for five shots using animal faces. For birds, an accuracy of 60.19% for one shot and 75.75% for five shots was obtained.
According to studies reviewed here, FSL can significantly boost the detection performance on small datasets. The popularity of FSL for smart farming applications is increasing rapidly due to the straightforward implementation yet highly promising results obtained in agricultural applications.
Statements
Author contributions
NR: writing—original draft and investigation. JT: writing—review and editing, supervision, and conceptualization. All authors contributed to the article and approved the submitted version.
Conflict of interest
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher’s note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
1
AfifiA.AlhumamA.AbdelwahabA. (2020). Convolutional neural network for automatic identification of plant diseases with limited data. Plants.10, 28. 10.3390/plants10010028
2
AlajajiD.AlhichriH. S.AmmourN.AlajlanN. (2020). “Few-shot learning for remote sensing scene classification,” in 2020 Mediterranean and Middle-East Geoscience and Remote Sensing Symposium (M2GARSS).IEEE. p. 81–84. 10.1109/M2GARSS47143.2020.9105154
3
Altae-TranH.RamsundarB.PappuA. S.PandeV. (2017). Low data drug discovery with one-shot learning. ACS Central Sci.3, 283–293. 10.1021/acscentsci.6b00367
4
ArgüesoD.PiconA.IrustaU.MedelaA.San-EmeterioM. G.BereciartuaA.et al. (2020). Few-Shot Learning approach for plant disease classification using images taken in the field. Comput. Elect. Agri.175, 105542. 10.1016/j.compag.2020.105542
5
AshtianiS. H. M.JavanmardiS.JahanbanifardM.MartynenkoA.VerbeekF. J. (2021). Detection of mulberry ripeness stages using deep learning models. IEEE Access.9, 100380–100394. 10.1109/ACCESS.2021.3096550
6
BrunelliD.AlbaneseA.d'AcuntoD.NardelloM. (2019). Energy neutral machine learning based iot device for pest detection in precision agriculture. IEEE Int. Things Magaz.2, 10–13. 10.1109/IOTM.0001.1900037
7
CaoZ.ZhangT.DiaoW.ZhangY.LyuX.FuK.et al. (2019). Meta-seg: A generalized meta-learning framework for multi-class few-shot semantic segmentation. IEEE Access.7, 166109–166121. 10.1109/ACCESS.2019.2953465
8
ChenW. Y.LiuY. C.KiraZ.WangY. C. F.HuangJ. B. (2019). A closer look at few-shot classification. arXiv.
9
ChenX.JiaS.XiangY. (2020). A review: knowledge reasoning over knowledge graph. Expert Syst. Appl.141, 112948. 10.1016/j.eswa.2019.112948
10
ChenY.LiuZ.XuH.DarrellT.WangX. (2021). “Meta-baseline: Exploring simple meta-learning for few-shot learning,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (Montreal, QC: IEEE), 9062–9071. 10.1109/ICCV48922.2021.00893
11
DongX.ShenJ. (2018). “Triplet loss in iamese network for object tracking,” in Proceedings of the European conference on computer vision (ECCV) (p. 459–474).
12
DuanR.LiD.TongQ.YangT.LiuX.LiuX. (2021). A survey of few-shot learning: an effective method for intrusion detection. Sec. Communi. Net.2021. 10.1155/2021/4259629
13
Fei-FeiL.FergusR.PeronaP. (2006). One-shot learning of object categories. IEEE Trans. Pattern Anal. Mach. Intell.28, 594–611. 10.1109/TPAMI.2006.79
14
FinkM. (2004). Object classification from a single example utilizing class relevance metrics. Adv. Neural Inf. Process. Syst.17.
15
GidarisS.KomodakisN. (2018). “Dynamic few-shot visual learning without forgetting,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (Salt Lake City, UT: IEEE), 4367–4375.
16
HospedalesT.AntoniouA.MicaelliP.StorkeyA. (2021). Meta-learning in neural networks: A survey. IEEE Trans. Pattern Anal. Mach. Intell.44, 5149–5169. 10.1109/TPAMI.2021.3079209
17
HuG.WuH.ZhangY.WanM. (2019). A low shot learning method for tea leaf's disease identification. Comput. Electron. Agric.163, 104852. 10.1016/j.compag.2019.104852
18
JadonS. (2020). “SSM-net for plants disease identification in low data regime,” in 2020 IEEE/ITU International Conference on Artificial Intelligence for Good (Geneva: IEEE), 158–163.
19
JaiswalA.BabuA. R.ZadehM. Z.BanerjeeD.MakedonF. (2020). A survey on contrastive self-supervised learning. Technologies.9, 2. 10.48550/arXiv.2011.00362
20
JanarthanS.ThuseethanS.RajasegararS.LyuQ.ZhengY.YearwoodJ. (2020). Deep metric learning based citrus disease classification with sparse data. IEEE Access.8, 162588–162600. 10.1109/ACCESS.2020.3021487
21
JiZ.ChaiX.YuY.PangY.ZhangZ. (2020). Improved prototypical networks for few-shot learning. Pattern Recognit. Lett.140, 81–87. 10.1016/j.patrec.2020.07.015
22
Jiménez-LunaJ.GrisoniF.SchneiderG. (2020). Drug discovery with explainable artificial intelligence. Nat. Machine Intelli.2, 573–584. 10.1038/s42256-020-00236-4
23
KaramiA.CrawfordM.DelpE. J. (2020). “Automatic plant counting and location based on a few-shot learning technique,” in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing.p. 5872–5886.
24
KasinathanT.SingarajuD.UyyalaS. R. (2021). “Insect classification and detection in field crops using modern machine learning techniques,” in Information Processing in Agriculturep. 446–457.
25
KhodakM.BalcanM. F. F.TalwalkarA. S. (2019). Adaptive gradient-based meta-learning methods. Adv. Neural Inf. Process. Syst.32. 10.48550/arXiv.1906.02717
26
KrizhevskyA.SutskeverI.HintonG. E. (2017). Imagenet classification with deep convolutional neural networks. Commun. ACM.60, 84–90. 10.1145/3065386
27
KuzuharaH.TakimotoH.SatoY.KanagawaA. (2020). “Insect Pest Detection and Identification Method Based on Deep Learning for Realizing a Pest Control System,” in 2020 59th Annual Conference of the Society of Instrument and Control Engineers of Japan (SICE). p. 709–714.
28
LakeB.SalakhutdinovR.GrossJ.TenenbaumJ. (2011). “One shot learning of simple visual concepts,” in Proceedings of the annual meeting of the cognitive science society.p. 33.
29
LakeB. M.SalakhutdinovR.TenenbaumJ. B. (2015). Human-level concept learning through probabilistic program induction. Science.350, 1332–1338. 10.1126/science.aab3050
30
LeeS. H.ChanC. S.MayoS. J.RemagninoP. (2017). How deep learning extracts and learns leaf features for plant classification. Pattern Recognit.71, 1–13. 10.1016/j.patcog.2017.05.015
31
LiX.PuF.YangR.GuiR.XuX. (2020). AMN: Attention metric network for one-shot remote sensing image scene classification. Remote Sens.12, 4046. 10.3390/rs12244046
32
LiY.ChaoX. (2021). Semi-supervised few-shot learning approach for plant diseases recognition. Plant Methods17, 1–10. 10.1186/s13007-021-00770-1
33
LiY.YangJ. (2020). Few-shot cotton pest recognition and terminal realization. Comput. Electron. Agric. 169, 105240. 10.1016/j.compag.2020.105240
34
LiY.YangJ. (2021). Meta-learning baselines and database for few-shot classification in agriculture. Comput. Electron. Agric. 182, 106055. 10.1016/j.compag.2021.106055
35
LiY.ZhangL.WeiW.ZhangY. (2020). “Deep self-supervised learning for few-shot hyperspectral image classification,” in IGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Symposium. IEEE. p. 501–504.
36
LiuJ.WangX. (2021). Plant diseases and pests detection based on deep learning: a review. Plant Methods.17, 1–18. 10.1186/s13007-021-00722-9
37
LiuL.WangR.XieC.YangP.WangF.SudirmanS.et al. (2019). PestNet: An end-to-end deep learning approach for large-scale multi-class pest detection and classification. IEEE Access.7, 45301–45312. 10.1109/ACCESS.2019.2909522
38
MelekhovI.KannalaJ.RahtuE. (2016). “Siamese network features for image matching,” in 2016 23rd international conference on pattern recognition (ICPR). IEEE. p. 378–383
39
MuthukumarN. (2021). “Few-shot learning text classification in federated environments,” in 2021 Smart Technologies, Communication and Robotics (STCR).p. 1-3.
40
NgH. F.LoJ. J.LinC. Y.TanH. K.ChuahJ. H.LeungK. H. (2022). “Fruit ripeness classification with few-shot learning,” in Proceedings of the 11th International Conference on Robotics, Vision, Signal Processing and Power Applications. Springer, Singapore. p. 715–720.
41
NuthalapatiS. V.TungaA. (2021). “Multi-domain few-shot learning and dataset for agricultural applications,” in Proceedings of the IEEE/CVF International Conference on Computer Vision.p. 1399–1408.
42
PanY.YaoT.LiY.WangY.NgoC. W.MeiT. (2019). “Transferrable prototypical networks for unsupervised domain adaptation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition.p. 2239–2247.
43
ParnamiA.LeeM. (2022). Learning from few examples: a summary of approaches to few-shot learning. arXiv. 10.48550/arXiv.2203.04291
44
PujaraJ.MiaoH.GetoorL.CohenW. (2013). Knowledge graph identification. In International semantic web conference. Berlin, Heidelberg: Springer. p. 542–557
45
ShaoL.ZhuF.LiX. (2014). Transfer learning for visual categorization: a survey. IEEE.26, 1019–1034. 10.1109/TNNLS.2014.2330900
46
SunC.ShrivastavaA.VondrickC.MurphyK.SukthankarR.SchmidC. (2018). “Actor-centric relation network,” in Proceedings of the European Conference on Computer Vision (ECCV)p. 318–334.
47
SungF.YangY.ZhangL.XiangT.TorrP. H.HospedalesT. M. (2018). “Learning to compare: Relation network for few-shot learning,” in Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 1199–1208).
48
TachaiN.YatoP.MuangpanT.SrijiranonK.EiamkanitchatN. (2021). “KaleCare: Smart Farm for Kale with Pests Detection System using Machine Learning,” in 2021 16th International Joint Symposium on Artificial Intelligence and Natural Language Processing (iSAI-NLP)p. 1–6.
49
TassisL. M.KrohlingR. A. (2022). Few-shot learning for biotic stress classification of coffee leaves. Artif. Intell. 6, 57–67. 10.1016/j.aiia.2022.04.001
50
TooE. C.LiY.KwaoP.NjukiS.MosomiM. E.KibetJ. (2019). Deep pruned nets for efficient image-based plants disease classification. J. Intell. Fuzzy Syst.37, 4003–4019. 10.3233/JIFS-190184
51
TorreyL.ShavlikJ. (2010). “Transfer learning,” in Handbook of Research on Machine Learning Applications and Trends: Algorithms, Methods, and Techniques (IGI Global), 242–264.
52
VinyalsO.BlundellC.LillicrapT.WierstraD. (2016). Matching networks for one shot learning. Adv. Neural Inf. Process. Syst.29.
53
WangB.WangD. (2019). Plant leaves classification: a few-shot learning method based on siamese network. IEEE Access.7, 151754–151763. 10.1109/ACCESS.2019.2947510
54
WangC.ZhouJ.ZhaoC.LiJ.TengG.WuH. (2021). Few-shot vegetable disease recognition model based on image text collaborative representation learning. Comp. Electronics Agri.184, 106098. 10.1016/j.compag.2021.106098
55
WangQ.MaoZ.WangB.GuoL. (2017). Knowledge graph embedding: a survey of approaches and applications. IEEE Trans. Knowl. Data Eng.29, 2724–2743.
56
WangY.YaoQ.KwokJ. T.NiL. M. (2020). Generalizing from a few examples: a survey on few-shot learning. ACM Comp. Surv.53, 1–34. 10.1145/3386252
57
WeissK.KhoshgoftaarT. M.WangD. (2016). A survey of transfer learning. J. Big data3, 1–40. 10.1186/s40537-016-0043-6
58
YangJ.GuoX.LiY.MarinelloF.ErcisliS.ZhangZ. (2022). A survey of few-shot learning in smart agriculture: developments, applications, and challenges. Plant Method.18, 1–12. 10.1186/s13007-022-00866-2
59
YangL.JinR. (2006). Distance metric learning: a comprehensive survey. Michigan State Uni.2, 4.
60
YaoY.ZhangY.NieW. (2020). “Pest Detection in Crop Images Based on OTSU Algorithm and Deep Convolutional Neural Network,” in 2020 International Conference on Virtual Reality and Intelligent Systems (ICVRIS). IEEE. p. 442–445
61
YavariN. (2020). Few-Shot Learning with Deep Neural Networks for Visual Quality Control: Evaluations on a Production Line.
62
YeH. J.HuH.ZhanD. C.ShaF. (2020). “Few-shot learning via embedding adaptation with set-to-set functions,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (Seattle, WA: IEEE), 8808–8817.
63
ZhaiX.OliverA.KolesnikovA.BeyerL. (2019). “S4l: Self-supervised semi-supervised learning,” in Proceedings of the IEEE/CVF International Conference on Computer Vision.p. 1476–1485.
64
ZhangH.KoniuszP.JianS.LiH.TorrP. H. (2021). “Rethinking class relations: Absolute-relative supervised and unsupervised few-shot learning,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (Nashville, TN: IEEE), 9432–9441.
65
ZhangS.ZhuJ.LiN. (2021). “Agricultural Pest Detection System Based on Machine Learning,” in 2021 IEEE 4th International Conference on Electronics Technology (ICET).IEEE. p. 1192-1196.
66
ZhongF.ChenZ.ZhangY.XiaF. (2020). Zero-and few-shot learning for diseases recognition of Citrus aurantium L. using conditional adversarial autoencoders. Comput Electron Agric.179, 105828. 10.1016/j.compag.2020.105828
67
ZhongQ.ChenL.QianY. (2020). “Few-shot learning for remote sensing image retrieval with maml,” in 2020 IEEE International Conference on Image Processing (ICIP). IEEE. p. 2446–2450.
Summary
Keywords
smart agriculture, object detection, computer vision, digital image processing, machine learning, few-shot learning
Citation
Ragu N and Teo J (2023) Object detection and classification using few-shot learning in smart agriculture: A scoping mini review. Front. Sustain. Food Syst. 6:1039299. doi: 10.3389/fsufs.2022.1039299
Received
27 September 2022
Accepted
19 December 2022
Published
18 January 2023
Volume
6 - 2022
Edited by
Muhammad Faisal Manzoor, Foshan University, China
Reviewed by
Seyed-Hassan Miraei Ashtiani, Ferdowsi University of Mashhad, Iran; Ranjana Roy Chowdhury, Indian Institute of Technology Ropar, India
Updates
Copyright
© 2023 Ragu and Teo.
This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Jason Teo ✉ jtwteo@ums.edu.my
This article was submitted to Sustainable Food Processing, a section of the journal Frontiers in Sustainable Food Systems
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.