Towards Transferable Pollution Detection Methods in Aquatic Environments Using Hyperspectral Technology

Abstract

This thesis presents a suite of transferable methodologies for environmental monitoring, focusing on detecting pollutants in aquatic ecosystems. Leveraging remote sensing with hyperspectral imaging (HSI) and artificial intelligence (AI), these approaches enable precise pollutant identification and adaptability across diverse scenarios. Thus, supporting systematic observation over extensive geographic regions or long-term datasets contributes to developing standardized solutions for monitoring and protecting Earth’s ecosystems. Aquatic pollution threatens biodiversity and ecosystem services, especially oil spills and plastic waste. Oil spills spread quickly across water surfaces, blocking sunlight and endangering aquatic biodiversity. Plastic waste accumulates in water bodies, breaking into microplastics that marine organisms ingest, disrupting food chains. Both pollutants are especially critical to monitor because they have profound, long-term impacts on ecosystems and food webs. Dye tracers such as rhodamine can be used as a proxy in oil spill simulations due to their similar dispersion behaviour in water. This aids in the development of detection and monitoring techniques for real oil spill events. However, there is a need for efficient and transferable aquatic pollution monitoring technologies. HSI is a powerful tool for identifying aquatic pollutants due to its rich spectral detail. However, the massive data it generates is costly and complex to process, posing key challenges. Dimensionality reduction techniques—such as spectral indices and band selection—can trim redundant data, reducing storage and computational demand while maintaining accuracy. Furthermore, the scarcity of labelled datasets hinders AI model training. Unsupervised learning approaches offer a promising solution, enabling models to extract meaningful patterns from non-labelled data, making HSI more adaptable across diverse environments. This thesis aims to pioneer efficient and transferable HSI methodologies for detecting and monitoring critical aquatic pollutants. It focuses on developing novel approaches that streamline data analysis and improve transferability across diverse environments. The research progresses from wellestablished, straightforward methods such as spectral indexes to cutting-edge AI methodologies to enhance pollutant detection. Each chapter focuses on different HSI technology needs, overcoming the challenges of data complexity, dimensionality, and the scarcity of labelled datasets. This creates a cohesive framework that addresses the demands of large-scale environmental monitoring with adaptable HSI methods. Chapter 2 reduces HSI data complexity by introducing the Normalized Difference O il I ndex ( NDOI), a n ew s pectral i ndex d esigned t o i mprove oil spill detection in coastal areas. The study compares the performance of several spectral indices utilizing images from multiple satellite and airborne sensors—AVIRIS, HICO, and MERIS—captured during the Deepwater Horizon disaster in the Gulf of Mexico in 2010. Traditional indices often misclassify other elements, such as suspended sediments, leading to inaccurate results. The NDOI avoids sand-in-suspension false positives, offering a m ore reliable response in coastal areas. NDOI is suitable for detecting oil spills thicker than 50 microns, with an average oil F1-score of 83%, and estimating its thickness and oil volume exceeding 90% accuracy. It also provides rapid detection of oil spills due to its simple calculation compared with other spectral indices and IA models, which is crucial for quick responses to environmental crises. This development directly contributes to optimizing the use of optical sensors for fast and efficient pollutant detection. Chapter 3 tackles the challenge of hyperspectral data’s high dimensionality by presenting a new dimensionality reduction methodology. The spectral band selection method identifies t he m ost r elevant b ands f or d etecting specific pollutants, with testing conducted on plastics and rhodamine. This minimises redundant or irrelevant bands, reducing computational cost and resource demands. This methodology has been applied to laboratory images and outdoor experiments, focusing on analyzing the impact of background effects on identifying target objects. The methodology successfully transferred influential spectral bands between datasets with 80-90% accuracy, indicating the potential for developing specialized sensors with these common bands to enable detection across various environments. However, the transfer of pre-trained classification models remains an area for further research, particularly regarding semitransparent objects or solutions influenced by background reflections in complex environments like optically shallow waters. Refined p ost-processing approaches suggest that model transfer could be feasible, potentially reducing the need for labelled data or in-situ validation, thus preserving resources and enabling a more generalizable classifier. Chapter 4 addresses the scarcity of labelled data in HSI and AI applications through the spectral loss function (Sl), which enhances HSI segmentation in unsupervised neural networks. The loss function is tested on HSI benchmark datasets, such as Pavia University, Salinas Valley, Indian Pines, and University of Houston, and a case study using an AVIRIS image from the Deepwater Horizon catastrophe. Sl was introduced in the currently best-performing unsupervised segmentation neural network, enhancing evaluation metrics performance by up to 6%. The proposed method also outperforms well-established techniques, such as spectral indices. For example, spectral indices rely on few spectral bands to produce a numerical value for each pixel, requiring an expert to establish a threshold for class determination. In contrast, the unsupervised neural network can directly assign different classes to varying oil thicknesses based on their complete spectral response, making them fully transferable across environments. Therefore, the unsupervised approach can generate ground-truth data, reducing manual labour. This is especially relevant for remote areas such as the open ocean, where manual labelling is challenging and resource-intensive. This contribution expands the scope of AI-driven detection techniques to operate without labelled datasets, thereby enhancing adaptability. This thesis concludes with a synthesis of the main insights from each chapter, a reflection on the implications and limitations identified, and suggestions for future research directions. The thesis successfully develops new transferable HSI methods that enhance the detection and monitoring of aquatic pollutants, overcoming critical knowledge gaps. Spectral indices, such as NDOI, provide a quick and efficient solution for rapid, low-resource decision-making but rely on manual thresholding. Band selection methods help identify critical spectral bands, which can improve model transfer and generalization across environments. Unsupervised methods complement these techniques by addressing non-labelled datasets, providing a foundation for large-scale monitoring. This research contributes to more efficient, adaptable, and transferable environmental monitoring technologies by addressing critical challenges related to data complexity, dimensionality reduction, and the scarcity of labelled datasets. The strengths and weaknesses of this suite of methods should be carefully considered to select the most appropriate approach based on each study’s specific characteristics. These advancements hold promise for designing next-generation sensors for UAVs and space missions, prioritizing data efficiency and precision. Moreover, the techniques directly apply to environmental management, including early spill detection, beach cleanup coordination, and supporting data-driven policies. Future work will focus on further automating hyperspectral monitoring techniques to minimize manual intervention. Efforts will be directed toward improving algorithm transferability by incorporating more variability in training data and advancing post-processing techniques. Additionally, the scalability of emerging tools, such as cloud computing, will be explored to improve efficient large-scale monitoring.