Deep learning-based water body extraction using high-resolution RGB-UAV imagery: a case study on Horseshoe Island, Antarctica

Abstract

Antarctica, which contains nearly 70% of the world's freshwater reserves, plays a key role in regulating global sea level and climate processes. Its coastal environments are highly dynamic and shaped by glacier calving, tidal forcing, and complex ocean-ice interactions, highlighting the need for accurate and high-resolution monitoring of coastal water bodies. While satellite remote sensing provides valuable long-term and regional-scale observations, its spatial resolution is often insufficient for capturing fine-scale coastal processes. In this context, unmanned aerial vehicles (UAVs) offer centimeter-level spatial resolution, enabling detailed environmental monitoring. This study evaluates deep learning-based water body extraction using high-resolution RGB-UAV imagery acquired over Horseshoe Island, Antarctica. A newly developed dataset comprising 287 RGB-UAV images was used to systematically compare recent six state-of-the-art semantic segmentation architectures: U-Net++, DeepLabv3+, MA-Net, SegFormer, ConvNeXt, and DINOv3. Model performance was assessed using the intersection over union (IoU) metric across different test regions. Among the evaluated models, MA-Net achieved the highest performance, with an IoU of 0.9513 and an overall accuracy of 0.9814. The model was further evaluated on several test areas with varying surface characteristics. Specifically, the model achieved an IoU of 0.9909 when tested over a large lake region. In addition, a preliminary evaluation of a lower-resolution Landsat image yielded an IoU of 0.9749, indicating promising performance across different spatial resolutions. Furthermore, on a detailed coastal area characterized by complex shoreline geometries and fragmented sea ice, the model yielded an IoU of 0.9611. These results indicate that the proposed framework can effectively delineate complex features under challenging Antarctic conditions and show potential as a practical and scalable approach for high-resolution coastal water body monitoring in polar environments using RGB-only data.