Bioinformatics in Agriculture and Ecology Using Few-Shots Learning From Field to Conservation

Bioinformatics in Agriculture and Ecology Using Few-Shots Learning From Field to Conservation

Jayashri Prashant Shinde (G.H. Raisoni College of Engineering and Management, Pune, India), Smitha Nayak (Muscat College, University of Stirling, Oman), Deepika Amol Ajalkar (G.H. Raisoni College of Engineering and Management, Pune, India), and Yogesh Kumar Sharma (Koneru Lakshmaiah Education Foundation, India)
DOI: 10.4018/979-8-3693-1822-5.ch002
OnDemand:
(Individual Chapters)
Available
$37.50
No Current Special Offers
TOTAL SAVINGS: $37.50

Abstract

The integration of bioinformatics with contemporary machine-learning algorithms is transforming sustainable practices and conservation activities in biology and agriculture. Plant disease identification is an area where few-shot learning (FSL) excels because of data scarcity. This study applies FSL to computational biology to tackle agricultural and environmental concerns. Bioinformatics has a significant influence on sustainable farming and research, according to the report. The chapter introduces few-shot learning, and shows how it may address the lack of labelled data in several disciplines. Case studies, including explanations, demonstrate the manner in which the FSL method is widely used in ecological surveillance, environmental programs, and crop supervisors. The essay discusses ethical issues around machine learning in ecological systems and agriculture, emphasizing open and responsible data methods.
Chapter Preview
Top

1. Introduction

A significant contributor to lower agricultural yields and quality is plant diseases. Computer vision using deep learning has made great strides in the last few years. Because of their affordable and readily available nature, image-based deep learning approaches to auto-diagnosis have caused significant alarm in the agricultural sector. Both farmers and the farm sector stand to gain substantially from the use of this technology. One problem with using deep-learning techniques is the data scarcity Singh, K., & Malhotra, D. 2023).

We all know that deep-learning techniques can't function without massive amounts of data. Unfortunately, there is a severe lack of plant-disease identification data in the current datasets. Developing comprehensive databases of plant disease photos, however, is a time-consuming and expert-dependent process. Data augmentation, FSL, transfer learning (TL), etc., are some of the methods that have been suggested to address data scarcity (Song Y. et al., 2023).

Recently, FSL has been suggested as a way to generalize to new categories using just a small number of examples. As a result, the pressing need for data has been successfully met. A subset of FSL is known as the metric-based approach. The basic idea is that features from samples in the same category tend to be near each other, whereas features from models in different varieties tend to be far apart. Training with a positive or negative sample pair, the Siamese network was the first representative study (Ahmed, S. (2022).

Afterwards, a series of classical networks were suggested. The idea of “seq2seq+attention” is used by matching networks to train the nearest neighbour classifier from end to end. The standard network develops to align with the class's proto-centre in the semantic environment using a small number of examples (Ge. Y. et al., 2023). A bidirectional network may learn about class relationships by merging the feature vectors of support and query samples. In order to determine how well the query sample matches the new classes, CoveMNet uses an embedding local covariance to extract second-order statistic information for each category. The central point of the grouping is the mean vector of the favourable samples in the naïve metric-based meta-baseline approach. It finds out how far away the centroid of new categories is from the query sample. The study of plant-disease identification has just recently begun to make use of FSL (Ge. Y. et al., 2023).

1.1 Problem Statement

A large-scale dataset is necessary for the image-based deep learning approach to plant disease diagnosis, but it shows promise. At the moment, using deep learning algorithms is hindered by the lack of data. For plant disease categories with limited sample sizes, the ability of few-shot learning to generalize to new categories with few samples is a huge boon. Two complex issues, however, arise in few-shot learning: (1) the number of shots used to extract features is small, and (2) it is not easy to generalize to new categories, particularly in a different area.

1.2 Motivation

Over the last decade, bioinformatics has grown in prominence as a technology to aid in the discovery of new biological information. The growth of the Human Genome Project has resulted in an enormous and fantastic expansion of biological knowledge. Capturing, managing, processing, analyzing, and interpreting data became more vital. Researchers may use bioinformatics with FSL algorithms to find a solution (Li W. et al., 2023). Dahiya et al. (2023) said this research is driven by agriculture's data collection restrictions and the data needs of machine learning algorithms. We use Few-Shot Learning to increase the computational prediction of the model's precision, especially with few labelled data. This supports environmentally friendly farming and surveillance of biodiversity by enhancing data utilisation and allowing more appropriate solutions.

Complete Chapter List

Search this Book:
Reset