Flood Inundation Mapping with Limited Observations Based on Physics-Aware Topography Constraint

Area of Science:

• Geographic Information Science
• Machine Learning
• Remote Sensing

Background:

• Spatial classification with limited observations is crucial for applications like flood mapping using partial geographic data.
• Existing methods struggle with datasets where most feature observations are missing, limiting their effectiveness.

Purpose of the Study:

• To extend a physics-aware structural constraint model for spatial classification to accommodate multi-modal feature distributions.
• To improve classification accuracy and robustness in scenarios with a high proportion of missing data.

Main Methods:

• Incorporation of physics-aware structural constraints into model representation.
• Development of efficient learning and inference algorithms.
• Extension to allow multi-modal feature distributions within classes.

Main Results:

• The proposed approach significantly outperforms baseline methods in classification accuracy for flood mapping.
• The multi-modal extension demonstrates greater robustness compared to the single-modal version.
• The solution is computationally efficient for large-scale datasets.

Conclusions:

• The developed method offers a robust and accurate solution for spatial classification with extensive missing observations.
• The multi-modal extension enhances the applicability of physics-aware models in complex geographical scenarios.
• The approach is computationally viable for practical, large-scale applications.