Resampling Techniques for Unbalanced Classification of Hyper/Multi-spectral Data in Agriculture Applications

Abstract

Developing machine learning algorithms for agriculture applications is becoming increasingly popular due to its applicability and ability to provide real-time information and decisions. Multispectral and hyperspectral imaging can provide more direct information about plants than other conventional means of sensing. Researchers conventionally use vegetation indices (VIs) derived from spectral information at specific bands to identify plant conditions. However, recent research shows that a machine learning (ML) approach which feed by large amount of information from M/HSI data is more effective than VIs. ML models usually need to be trained with sufficient and balanced training data. However, obtaining enough samples in certain situations can be challenging. Training machine learning models on imbalanced datasets, where some classes have significantly more samples than others, can degrade model performance of ML moodel. The easiest way to solve the imbalance data problem is by manipulating the number of samples in the training set (Resampling). Synthetic Minority Oversampling Technique (SMOTE) is one of the most popular oversampling approaches. However, this approach is considered less effective in data with hundreds of attributes or features, such as hyperspectral data, because defining neighborhood samples accurately becomes more difficult as data dimensions increase. In order to solve this issue, we evaluated several oversampling techniques to determine their suitability for different dataset characteristics. Various distance functions that are appropriate for high-dimensional data were also tested. In addition, we enhanced the algorithm with ensemble learning using diverse sampling rates to obtain a range of perspectives from the minority class. Experimental results on plant hyperspectral datasets from the University of Manchester demonstrated that employing the Manhattan distance function in oversampling techniques significantly improves performance compared to using the Euclidean distance function. Through this modification, we were able to improve the Matthew Correlation Coefficient (MCC) score by 0.9%, and the algorithm with the proposed ensemble learning, the MCC score is boosted to 1%. The reported improvement may seem modest at first glance. However, in precision agriculture, such enhancements are significant. Even small increases in accuracy can lead to substantial practical benefits when monitoring large-scale agricultural operations. For instance, a 1-2% improvement in detecting stressed or diseased crops on a large farm can greatly enhance the precision of interventions, optimizing resource usage and preventing disease spread. Another issue on conventional oversampling technique is although decision-based oversampling methods are reliable in overcoming imbalance problems in ambiguous samples, which are more likely to be misclassified than other samples. However, these methods sacrifice the ability to detect the class of ambiguous samples in the majority class. Here, we propose a modification to SVM-SMOTE that considers the presence of majority samples and includes a mechanism to determine the location for generating synthetic samples relative to the support vector of the majority class. This approach ensures that the generated samples minimally interference the decision boundary of the majority class. In our tests with the Bonn-Spec dataset, we can improved the MCC score by up to 3% compared to standard SVM-SMOTE. Besides conventional oversampling, GAN (Generative Adversarial Network) has potential to generate realistic data and overcome the problem of data imbalance. However, the synthetic samples generated by GAN may not always positively affect the quality of decision boundary to improve the classifier performance. This is because data may be generated in less effective areas. To address this challenge, we designed a framework by modifying GAN and combining it with SMOTE as a guide to generate new data. A filter mechanism was designed to ensure that the newly generated data has a positive effect on data imbalance. This approach significantly improved classification results on the Indian Pines dataset, with a 3% MCC improvement over SMOTE and a 0.7% improvement over the original GAN. The report indicates that the improvement demonstrates that the proposed GAN-based oversampling technique can generate high-quality synthetic minority data, which positively contributes to enhance the performance of machine learning models. The significance of this research lies in its potential to produce robust machine learning models even with limited data. Moreover, it reduces the need to incur high costs to obtain rare agriculture specimens, as the proposed GAN model can automatically generate high-quality synthetic data.

Date of Award	8 Jan 2025
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Hujun Yin (Main Supervisor) & Bruce Grieve (Co Supervisor)