Data Augmentation Techniques for Machine Learning Applied to Optical Spectroscopy Datasets in Agrifood Applications:...

Area of Science:

• Agricultural Science
• Data Science
• Spectroscopy

Background:

• Machine learning (ML) and deep learning (DL) excel in various tasks but require extensive data.
• Agrifood applications often face data scarcity due to collection constraints and imbalanced datasets.
• Optical spectroscopy generates valuable data for the agrifood sector, but dataset limitations hinder ML/DL model performance.

Purpose of the Study:

• To review data augmentation (DA) techniques for optical spectroscopy datasets in the agrifood industry.
• To explore how DA addresses data scarcity and imbalanced datasets in this domain.
• To present both simple and advanced DA methods, including generative adversarial networks (GANs).

Main Methods:

• Review of existing literature on data augmentation applied to agrifood optical spectroscopy data.
• Categorization of DA techniques from simple sample duplication to complex deep learning models.
• Analysis of generative adversarial networks (GANs) and semi-supervised generative adversarial networks (SGANs) for synthetic data generation.

Main Results:

• Data augmentation effectively expands limited optical spectroscopy datasets for agrifood applications.
• Techniques range from basic data replication to sophisticated generative models like GANs and SGANs.
• DA helps mitigate issues of imbalanced datasets common in agrifood data collection.

Conclusions:

• Data augmentation is crucial for building robust ML/DL models using optical spectroscopy in the agrifood industry.
• Advanced techniques like GANs offer powerful solutions for generating synthetic data to overcome limitations.
• The application of DA in this field promises to improve analytical capabilities and decision-making.