IEEE International Geoscience and Remote Sensing Symposium 2023 — First Author & Session Chair

I delivered a first author oral presentation at IGARSS 2023 and chaired the session “Image Analysis for Remote Sensing of Water Bodies” . The talk presented our work titled “Water Physics Aware Semantic Segmentation Through Texture-Biased U-Net Architectures”, which investigates how the physical properties of water can be exploited to improve semantic segmentation performance in aerial imagery.

Paper:
“Water Physics Aware Semantic Segmentation Through Texture-Biased U-Net Architectures”
IEEE IGARSS 2023

The presentation covered:

• Physical properties of water that influence its appearance, including surface tension, cohesion, adhesion, and vibrational colour origins.
• Why colour is unreliable for water segmentation due to reflections, impurities, scattering, and illumination changes.
• Motivation for texture-biased segmentation based on water’s consistent surface texture.
• Introduction of two texture-biased architectures:
  • GUNet — U-Net with a Gabor-implemented convolutional first layer and Mix Pooling.
  • GMACUNet — MACUNet with the same texture-biased modifications.
• Construction of a diverse water dataset with varying seasons, lighting, atmospheric conditions, snow/ice, and reflections.
• Evaluation across three benchmark aerial datasets: Landcover.AI, WHDLD, and UAVid.

Key findings:

• Texture-biased architectures outperform standard U-Net and MACUNet across all datasets.
• Gabor-based convolutional layers improve robustness to shadows, canopy occlusion, and light variation.
• The proposed models retrieve scene information hidden beneath canopy and shadows more effectively.
• The approach generalises strongly beyond water segmentation to multi-class aerial scene segmentation.

Session role:

• Chaired the IGARSS 2023 session “Image Analysis for Remote Sensing of Water Bodies” in which the paper was presented, coordinating speakers and discussion.