Related Experiment Video
Updated: Jun 4, 2026
09:19
Measuring the Structure, Composition, and Change of Underwater Environments with Large-area Imaging
Published on: April 18, 2025
Deep learning-based water body extraction using high-resolution RGB-UAV imagery: a case study on Horseshoe Island,
Mahmut Oğuz Selbesoğlu1, Bahadır Kulavuz2, Ceyda Hızal2
1Faculty of Civil Engineering, Geomatics Engineering, İstanbul Technical University, Istanbul, Turkey. selbesoglu@itu.edu.tr.
Environmental Monitoring and Assessment
|June 3, 2026
Summary
Deep learning models accurately map Antarctic coastal waters using high-resolution drone imagery. MA-Net excels in delineating water bodies, crucial for monitoring these dynamic polar environments.
Area of Science:
- Environmental monitoring
- Remote sensing
- Artificial intelligence in Earth sciences
Background:
- Antarctica's coastal environments are critical for global sea level and climate regulation.
- High-resolution monitoring of Antarctic water bodies is essential due to dynamic coastal processes.
- Satellite remote sensing lacks the spatial resolution needed for fine-scale Antarctic coastal analysis.
Purpose of the Study:
- To evaluate deep learning models for water body extraction from high-resolution Unmanned Aerial Vehicle (UAV) imagery in Antarctica.
- To compare the performance of six state-of-the-art semantic segmentation architectures.
- To assess the potential of RGB-only UAV data for high-resolution polar coastal monitoring.
Main Methods:
- Development of a new dataset with 287 RGB-UAV images from Horseshoe Island, Antarctica.
- Systematic comparison of six semantic segmentation models: U-Net++, DeepLabv3+, MA-Net, SegFormer, ConvNeXt, and DINOv3.
- Performance evaluation using the Intersection over Union (IoU) metric across diverse Antarctic coastal regions.
Main Results:
- MA-Net achieved the highest performance, with an IoU of 0.9513 and overall accuracy of 0.9814.
- MA-Net demonstrated excellent performance on various surfaces, including a large lake (IoU 0.9909) and complex coastlines with sea ice (IoU 0.9611).
- Preliminary evaluation on lower-resolution Landsat imagery showed promising IoU of 0.9749, indicating cross-resolution applicability.
Conclusions:
- Deep learning, particularly MA-Net, effectively delineates Antarctic coastal water bodies using high-resolution UAV imagery.
- The proposed framework shows potential for practical and scalable high-resolution monitoring in challenging polar environments.
- RGB-only UAV data offers a viable solution for detailed coastal water body mapping in polar regions.