Degree

Doctor of Philosophy (PhD)

Department

The Department of Geography and Anthropology

Document Type

Dissertation

Abstract

Wetlands, as a crucial component of the Earth's ecosystem, play a vital role in maintaining ecological balance and preserving biodiversity. However, wetlands are currently facing severe challenges, as both natural and human-induced factors are contributing to their global degradation. Gaining a deep understanding of the species composition and biodiversity within wetland-covered areas and accurately analyzing their changing trends not only helps assess the current and future state of wetlands but also provides strong support for the formulation of scientifically sound wetland conservation strategies. In recent years, the rapid development of geospatial intelligence and remote sensing technologies has brought new opportunities for wetland conservation efforts. Using the Louisiana wetland as the study area, this paper aims to explore cutting-edge AI and ML methods and introduce these methods to wetland land cover classification. To fully examine the ML/AI ability supporting the wetland classification, this paper covers different models with different input data such as hyper-spectral, LiDAR, and satellite Landsat Data. By introducing and comparing some cutting-edge and various AI/ML models and components that focus on different tasks, this paper aims to experiment and introduce AI/ML to wetland-related studies that could benefit the wider community. The works are summarized below. First, this paper proposes the Unet-processed Mobile Graph Convolution network (UMGC) for analyzing hyperspectral images (HSI). UMGC employs Mobile Graph Convolution (MGC) to construct dynamic graphs more efficiently than traditional KNN-based methods. This approach is particularly suited for multi-band HSI datasets, combining computational efficiency with the - 2 - flexibility to adapt connections across varying input images. Then, this paper proposes two new networks based on GNNs and Mamba, leveraging multi-source HSI and LiDAR data to improve classification results. Three components—Gated Recurrent Units (GRUs), Vision Transformers (ViTs), and U-Net—were selected to enhance performance. By integrating GRUs, ViTs, and U-Net within GNN and Mamba frameworks, our proposed networks aim to harness the strengths of these components to address the challenges in multi-source HSI and LiDAR data classification. Then, this paper also examined the models focused on Landsat data, which are more suitable for long-term and wide-scale studies. This paper proposes a novel wetland classification method, Attention-based Light-weight Conditionally parameterized Convolution (ALCC) network, suitable for large-scale and long-term series. The model is used to produce a 20-year time series of wetland land cover for Louisiana. Based on the new Louisiana time series, this paper then explored the prediction model. Based on the time series, this paper proposes a prediction model based on Unet and LSTM with Cross Entropy Loss (Unet-LSTM-CEL).

Date

3-31-2025

Committee Chair

Wang, Lei

Download

Available for download on Tuesday, March 31, 2026

Share

COinS