Wetland characterization through multi-sensor satellite data and advanced AI models

Jafarzadeh, Hamid (2025) Wetland characterization through multi-sensor satellite data and advanced AI models. Doctoral (PhD) thesis, Memorial University of Newfoundland.

[English] PDF - Accepted Version Available under License - The author retains copyright ownership and moral rights in this thesis. Neither the thesis nor substantial extracts from it may be printed or otherwise reproduced without the author's permission. Download (290MB)

Abstract

Wetlands are generally defined as areas that are inundated or saturated with water for at least part of the year, though definitions can vary significantly across different scientific disciplines. Recognized as some of Earth's most valuable and productive ecosystems, wetlands provide a range of essential ecological services. These include regulating global climate through carbon storage, naturally purifying water by filtering pollutants, mitigating ood and drought impacts, protecting shorelines from erosion, conserving soil, and supporting biodiversity by serving as critical habitats for diverse wildlife species. Wetlands also offer spaces for recreation, cultural appreciation, and aesthetic enjoyment, underscoring their multifunctional importance for environmental health and human well-being. They are dynamic ecosystems characterized by complex and uctuating vegetation, water, and soil interactions. Traditional monitoring methods rely on in-situ sampling of vegetation, hydrology, and soil, which are labor-intensive, time-consuming, and challenging to scale across vast or inaccessible regions. In contrast, remote sensing (RS) technology, coupled with advanced artificial intelligence (AI) techniques, offers an efficient, scalable, and non-intrusive approach for assessing wetland health, spatial extent, and ecological status. Through the integration of multisource - multispectral, Synthetic Aperture Radar (SAR), and Light Detection and Ranging (LiDAR) data - RS enables high-resolution monitoring across long-term periods and expansive spatial scales, allowing for efficient mapping and characterization of different wetland classes. These tools are crucial in quantifying wetland changes, identifying anthropogenic impacts, and understanding seasonal ecological variations in wetlands, thereby facilitating data-driven conservation and sustainable management efforts. This thesis focuses on improving wetland mapping and characterization by introducing novel methodologies that utilize advanced RS technologies and multisource satellite data. First, it presents a comprehensive literature review on wetland monitoring, focusing on the integration of RS and machine learning (ML) techniques. In the context of rapidly changing wetland landscapes, the synergy between multisource RS data and ML algorithms offers an opportunity to enhance decision-making support for wetland management. The review provides guidance on selecting suitable ML algorithms and RS data types for detailed wetland monitoring. A second key contribution of this work is the development of a novel deep learning (DL) framework to address the challenge of limited sample data, advancing the potential of RS techniques for wetland mapping. Emphasizing the integration of SAR and optical data, the framework utilizes multimodal data sources and applies Graph Convolutional Networks (GCN) and Convolutional Neural Networks (CNN). The experimental results demonstrate significantly improved classification accuracy, particularly in mapping complex ecosystems, compared to single-stream approaches. Thirdly, this research highlights the importance of RADARSAT Constellation Mission (RCM) compact polarimetric SAR (CP-SAR) data in wetland studies. A novel framework is developed to extract robust features for wetland mapping, showcasing notable improvements in wetland classification performance compared to existing models. Fourthly, a new SAR decomposition technique is introduced, designed specifically for CP-SAR data from RCM imagery. This technique refines the analysis of wetland structures and compositions. Additionally, a unique descriptor termed compact polarimetry signature (CPS) is proposed based on time-series RCM data, enabling a detailed investigation of wetland phenological cycles and associated SAR backscatter variations during active growing seasons. The results demonstrate the superior effectiveness of these techniques in characterizing complex wetland landscapes, surpassing the accuracy and detail of traditional single-date approaches. Finally, this work evaluates the added value of incorporating diverse data sources, including optical, SAR, and LiDAR data, into wetland mapping. Using the Google Earth Engine (GEE) platform, two main objectives are pursued: (1) integrating Global Ecosystem Dynamics Investigation (GEDI) LiDAR footprint heights with multisource datasets to generate vegetation canopy height (VCH) maps, and (2) enhancing wetland mapping by using VCH as a predictive input. Results highlight the critical role of VCH derived from GEDI samples in improving wetland classification accuracy, offering a vertical vegetation profile that enriches understanding of wetland ecosystems. This thesis introduces innovative methodologies that pave the way for practical and scalable solutions in wetland mapping. The approaches outlined not only enhance our ability to map complex wetland systems in Canada but also provide adaptable strategies applicable to ecologically similar wetlands around the world.

Actions (login required)

View Item