Unsupervised surface water mapping with airborne LiDAR data by leveraging physical properties of water

Comprehensive mapping of surface water, especially smaller bodies of water (<1 ha), remains challenging due to the lack of robust and scalable extraction methods. Traditional methods require the use of either training procedures or the repetitive tuning of site-specific parameters, which present hurdles to automated mapping and introduce biases tied to training data and parameters. The dependence on water’s reflectance properties, including LiDAR intensity, further complicates the issue, as higher-resolution images inherently introduce increased noise. In response to these challenges, we propose a unique, unsupervised method that focuses on the geometric characteristics of water instead of its variable reflectance properties. Unlike existing approaches, our method relies exclusively on 3D coordinate observations from airborne LiDAR data, taking advantage of the presumption that connected surface water remains flat due to surface tension. Leveraging this physical constraint and spatial connectivity, our method precisely extracts water bodies of diverse sizes and reflectance without the need for training procedures or intensive parameter tuning. Notably, by relying solely on 3D coordinate observations, our approach significantly facilitates the fully automated generation of comprehensive 3D topographical maps of both water and terrain, eliminating the need for human intervention or supplementary optical imagery. We validated the robustness and scalability of this method across diverse terrains, including urban, coastal, and mountainous areas. Overall, the proposed method achieved a 11% higher accuracy, measured by intersection over union, compared to the highly competitive NDWI-based method. Moreover, it proved its effectiveness in both accuracy and scalability compared to supervised machine learning and deep learning methods.