Original papersMeta-learning baselines and database for few-shot classification in agriculture
Introduction
Automatic recognition of plant leaf diseases and crop pests is an essential issue in agricultural production, guaranteeing crops' yield and quality (Sethy et al., 2020, Thenmozhi and Reddy, 2019, Li and Chao, 2020). Specifically, early identification of plant diseases and pests is necessary to monitor and warn the crop growth situation, which is beneficial for farming management. However, experts or experienced farmers' manual observation is still the primary approach in many countries and areas, which is inefficient and highly empirical. Thus, the automatic classification of crop pests and leaf diseases is significant in the agricultural field, attracting several researches (Too et al., 2019a, Too et al., 2019b, Thenmozhi and Reddy, 2017).
Many researchers used deep learning to solve the classification problem in previous works and achieved high performances (Goluguri et al., 2020, Kamilaris and Prenafeta-Boldú, 2018, Trong et al., 2020, Lu et al., 2017). As known, deep learning is an essential branch of machine learning, including a vast number of trainable parameters in deep layers. To achieve a good performance and overcome the overfitting problem, the deep learning model relies on large amounts of data to train, called data-driven learning. However, in the real world, the data distribution in various fields is long-tailed (Yang et al., 2020a), which means it is hard or expensive to collect so many large-scale datasets for different deep learning applications. On the other hand, humans can quickly learn and migrate from just a few samples, making us think about whether learning from massive amounts of data is the desired intelligence. Hence, learning from few data to classify is a meaningful and promising study in practical applications due to the low cost of a few samples of data, also called few-shot classification.
At present, there are mainly three schemes to deal with the few-shot classification problems: data augmentation, transfer learning, and meta-learning. Data augmentation is an intuitive solution to generate more new instances or features, utilizing the image rotation and scale, mix-up, oversampling, and other related techniques (Yang et al., 2020b). Transfer learning aims to transfer the knowledge between the source domain and the target domain. It is assumed that there is sufficient data in the source domain for training, and then the trained network will be fine-tuned by a few samples in the target domain to maintain a good performance (Zhuang et al., 2020). Meta-learning, also called learn to learn, is a task-driven method proposed for the few-shot problems. During the task iteration, the model learns to learn from a few samples to complete the few-shot classification task (Snell et al., 2017, Sung et al., 2018).
In the agricultural field, there has emerged a handful of frontier research on few-shot classification (Hu et al., 2019, Argüeso et al., 2020, Li and Yang, 2020). Hu et al. (2019) used the conditional deep convolutional generative adversarial network (C-DCGAN) to generate augmented images of tea leaf diseases. The expanded data were used to train the VGG16 model with average identification accuracy of 90%. This work is using the data augmentation technique to solve the few-shot tea leaf diseases classification. Argüeso et al. (2020) split the PlantVillage dataset into a source (32 classes) and a target (6 classes) domain. A general-purpose CNN to learn to extract general plant leaf characteristics was trained on the source domain and transferred to the new target domain. The testing accuracy was above 90% using 80 images per class, called 80-shot. This work is utilizing the transfer learning technique to solve the few-shot plant diseases classification. Li and Yang (2020) prepared the training set as the triplets' format and adopted the triplet loss function to train a CNN feature extractor. A few samples of origin data can be combined to form many training triples to train the CNN network based on distance metric comparison. The testing accuracy was 95.4% and 96.2% on two different datasets. Generally, this work uses the transfer learning technique to solve the few-shot crop pests classification and focuses on the hardware realization.
Although the above works made positive attempts, there is still a lot of research space for the agricultural field's few-shot classification. For instance, the mentioned references all focused on the single-domain classification, that is to say, either pests or plants. But in fact, the cross-domain classification is more challenging while not given enough attention. The agricultural research community needs a comprehensive database that can perform both single-domain and cross-domain analysis. Also, the existing works have not involved the important meta-learning paradigm. As a typical solution of few-shot classification problems, there are significant differences between meta-learning and deep learning. It is necessary to introduce the meta-learning paradigm into the few-shot classification studies in the agricultural field from our perspective.
In this paper, we collected samples from publicly available resources to assemble a comprehensive database for few-shot classification, covering both pests and plants. We carried out the first work (to the best of our knowledge) to adopt the meta-learning paradigm to solve the agricultural field's few-shot classification problems. Extensive experiments were carried out to establish the baselines of average testing accuracy and then explore the effect of domain shift and meta-learning parameters, e.g., N-way, K-shot. Finally, we summarized and explained each few-shot factor's effect laws and provided some ideas on the future work to achieve further improvement.
The contributions of this work are three-fold:
- (1)
We collected samples from publicly available resources to assemble a balanced dataset for the few-shot classification, containing pests and plants to analyze the single domain and cross-domain.
- (2)
We carried out the first work of task-driven meta-leaning few-shot classification in the agricultural field, and 36 groups of comparison experiments were performed to establish average accuracy baselines.
- (3)
We explored the effect of N-way, K-shot, and domain shift on the performance of few-shot classification based on extensive experiments. The influence laws of each factor are summarized and explained in detail.
Section snippets
Materials
The used dataset should include both pests and plants to perform single-domain and cross-domain few-shot classification research. The number of samples per category can be relatively small. The samples are collected from publicly available resources, e.g., the pest samples are partly from the shared database in Li et al. (2020b), while the plant samples are partly from the PlantVillage (https://plantvillage.psu.edu). There are 20 categories (total 6000 samples) in this new balanced dataset,
Meta-learning for few-shot problem of ‘N-way K-shot’
A most typical problem of few-shot classification is N-way K-shot: There are N categories and K samples per category available for training or learning. The model is wished to distinguish the N classes and tested by several query samples from each category. Generally, K is 1, 5, 10; that is why called few-shot.
Meta-learning regards the N-way K-shot problem as a task unit in the target set. Unlike the data-driven deep learning methods, meta-learning is a task-driven approach, which learns to
Results
Extensive comparison experiments on the few-shot classification were carried out, according to the N-way K-shot configuration. The experimental hardware resource is the NVIDIA TITAN Xp with 12 GB memory. The software environment is the Jupyter Notebook with libraries of Tensorflow (version 1.12.0), Numpy (version 1.19.2), and OpenCV (version 4.1).
Meta-learning few-shot classification is a task-driven learning paradigm, whose tasks in the meta-train and meta-test stages are all prepared as N-way
Discussion
Look through the entire study, we want to discuss this work from the following four aspects: motivation, contributions, findings, and future work.
Motivation: Learning from a few samples to classify is a significant and promising study to alleviate the emerging drawback of deep learning: the high cost of collecting and annotating required large-scale datasets. Through the literature research, although there have been a handful of few-shot studies in the agricultural field, none of them has
Conclusion
Learning from a few samples to automatically recognize the pests or plants is difficult but promising to protect the agricultural yield and quality with a low cost of data. We introduced an intuitive task-driven learning scheme, namely meta-learning few-shot classification, which is meta-trained in the source set by mimicking testing tasks in the target set. A balanced database covering pests and plants was collected from the publicly available resources. Through literature research, to our
CRediT authorship contribution statement
Yang Li: Conceptualization, Methodology, Software, Writing - original draft. Jiachen Yang: Supervision, Project administration, Writing - review & editing.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgements
This work was partially supported by the National Natural Science Foundation of China (No. 31860333 and No. 61871283), the Foundation of Pre-Research on Equipment of China (No.61400010304), and the Major Civil-Military Integration Project in Tianjin, China (No.18ZXJMTG00170).
References (29)
- et al.
Few-Shot Learning approach for plant disease classification using images taken in the field
Comput. Electron. Agric.
(2020) - et al.
Towards weeds identification assistance through transfer learning
Comput. Electron. Agric.
(2020) - et al.
A low shot learning method for tea leaf’s disease identification
Comput. Electron. Agric.
(2019) - et al.
Deep learning in agriculture: a survey
Comput. Electron. Agric.
(2018) - et al.
Do we really need deep CNN for plant diseases identification?
Comput. Electron. Agric.
(2020) - et al.
Crop pest recognition in natural scenes using convolutional neural networks
Comput. Electron. Agric.
(2020) - et al.
Few-shot cotton pest recognition and terminal realization
Comput. Electron. Agric.
(2020) - et al.
Identification of rice diseases using deep convolutional neural networks
Neurocomputing
(2017) - et al.
A survey of public datasets for computer vision tasks in precision agriculture
Comput. Electron. Agric.
(2020) - et al.
Deep feature based rice leaf disease identification using support vector machine
Comput. Electron. Agric.
(2020)
Crop pest classification based on deep convolutional neural network and transfer learning
Comput. Electron. Agric.
A comparative study of fine-tuning deep learning models for plant disease identification
Comput. Electron. Agric.
A two-stage approach to few-shot learning for image recognition
IEEE Trans. Image Process.
Cited by (116)
Few-shot classification for sensor anomalies with limited samples
2024, Journal of Infrastructure Intelligence and ResiliencePlant disease recognition in a low data scenario using few-shot learning
2024, Computers and Electronics in AgricultureReinforcement learning based edge computing in B5G
2024, Digital Communications and NetworksCitation Excerpt :Compared with the traditional algorithm, DRL has the characteristics of strong expansibility and robustness to local observation [15], which can effectively solve the problem of task allocation in B5G communication. DRL combines the perception ability of Deep Learning (DL) [16] with the decision-making ability of Reinforcement Learning (RL), which is a kind of artificial intelligence method closer to human thinking mode. In DL methods, features of different inputs are extracted through deep neural networks [17].
Deep metric learning framework combined with Gramian angular difference field image generation for Raman spectra classification based on a handheld Raman spectrometer
2023, Spectrochimica Acta - Part A: Molecular and Biomolecular SpectroscopyTen deep learning techniques to address small data problems with remote sensing
2023, International Journal of Applied Earth Observation and GeoinformationSpeech emotion recognition based on meta-transfer learning with domain adaption
2023, Applied Soft Computing