ANALYZING BIOGEOGRAPHICAL TRENDS OF BENTHIC HABITATS IN GUAM USING WORLDVIEW-2/3 IMAGERY INTEGRATED WITH ADVANCED DEEP LEARNING
Date
2025-05
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
University of Delaware
Abstract
Benthic habitatss—ecological communities inhabiting seafloor environments—
such as seagrass meadows and coral reefs, are vital components of coastal ecosystems
but face increasing threats from anthropogenic and climatic stressors. Traditional
remote sensing methods for mapping these habitats often struggle with fine-scale het erogeneity, label scarcity, and computational inefficiency. This study addresses these
challenges by integrating high-resolution WorldView-2/3 multispectral imagery with an
optimized deep learning framework to analyze biogeographical trends in the ecologically
distinct coastal zones of Guam. We propose Tiny-UNet—a lightweight U-Net-based
architecture tailored for benthic habitat segmentation, incorporating channel-wise at tention mechanisms and bilinear upsampling to enhance spectral-spatial feature extrac tion while minimizing computational complexity. The model was trained on a dataset
of nine labeled images that span two regions: Angana Bay and Manell Geus, annotated
with seven habitat classes. Strategic patch-based preprocessing and deterministic data
augmentation are applied to mitigate class imbalance and environmental variability,
such as sunglint, radiometric inconsistency, and clouds. We conducted experiments
and evaluated the performance of our proposed solution, which exhibits outstanding
performance for segmenting and classifying each of seven benthic classes, as well as in
its broader adaptability, when compared to the original U-Net design. The efficiency of
our framework (214K parameters vs 31M in the original U-Net) and the preservation
of the boundaries underscore its potential for scalable coastal monitoring. However,
persistent challenges in transitional zones—driven by environmental noise and spectral
ambiguities—require hybrid physics-AI models to disentangle overlapping signatures, active learning for label efficiency, and multi-sensor fusion to address dynamic condi tions. Integrating synthetic data generation and transformer architectures could fur ther mitigate class imbalance and sensor artifacts, balancing computational efficiency, model generalizability, and ecological precision for scalable coastal monitoring. This
work advances AI-driven marine conservation by balancing computational efficiency with ecological precision, o↵ering a pathway for adaptive management in data-limited
coastal ecosystems.