Jadwal Sholat

Memuat jadwal sholat…

Pertanian & Lingkungan editorial

Open AccessOA2022

A Survey of Few-Shot Learning in Smart Agriculture: Developments, Applications, and Challenges

This survey reviews few-shot learning methods—data augmentation, metric learning, external memory, and parameter optimization—that enable deep learning in agriculture with minimal labeled data, and discusses datasets, applications, and future challenges.
Jiachen Yang; Xiaolan Guo; Yang Li; F. Marinello; S. Ercişli; Zhuo Zhang· Plant Methods· 2022· DOI 10.1186/s13007-022-00866-2

The core problem

Deep learning has achieved remarkable success in agriculture and plant science, but its performance hinges on massive labeled datasets. In plant science and biology, acquiring large amounts of labeled data is often impractical due to cost, time, and expertise required. Few-shot learning (FSL) addresses this by mimicking human rapid learning: it can learn a new task from only a few labeled samples, drastically reducing time and financial resources. This review comprehensively examines FSL in smart agriculture, covering its definition, four main methodological categories, publicly available datasets, various applications, and future challenges.

Innovation

The survey categorizes advanced few-shot learning methods into four main approaches, each addressing the overfitting problem from different perspectives:

1. **Data augmentation**: Synthesizes additional training samples to enlarge the support set.
2. **Metric learning**: Learns a distance metric to compare query samples with few labeled examples.
3. **External memory**: Augments the model with memory modules to store and retrieve information from limited data.
4. **Parameter optimization**: Meta-learns initialization or optimization strategies for fast adaptation.

These methods are analyzed in the context of agricultural tasks, where data scarcity is prevalent.

Introduction
Deep learning has achieved remarkable success in agriculture and plant science, but its performance hinges on massive labeled datasets. In plant science and biology, acquiring large amounts of labeled data is often impractical due to cost, time, and expertise required. Few-shot learning (FSL) addresses this by mimicking human rapid learning: it can learn a new task from only a few labeled samples, drastically reducing time and financial resources. This review comprehensively examines FSL in smart agriculture, covering its definition, four main methodological categories, publicly available datasets, various applications, and future challenges.

Why it matters

The survey discusses challenges for future development of FSL in smart agriculture. These include the need for standardized benchmarks, domain adaptation across different crops and environments, integration with other AI techniques, and the interpretability of few-shot models. The authors emphasize that while FSL reduces the labeling burden, careful design and validation are essential for real-world deployment. The review concludes by outlining promising research directions to overcome these challenges.

Who should read this

Pembaca Research Digest dan praktisi rumpun ini.

Opening member content…