POSTER SESSION 2
10:30–12:30, Tuesday, September 15
Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance
ABSTRACT 1188 | POSTER 122
BATHYUNET++: A CENTER-FOCUSED RECEPTIVE-FIELD NETWORK FOR HIGH-RESOLUTION BATHYMETRY MAPPING FROM SUPERDOVE IMAGERY
High-spatial-resolution (HR) bathymetry is crucial for ensuring navigation safety and effective environmental management in shallow waters. However, traditional bathymetry retrieval from medium-resolution satellites suffers from large uncertainties in the presence of complex substrates and strong spatial heterogeneity. HR sensors, such as PlanetScope SuperDove (~ 3.7 m), can better resolve fine-scale features but often suffer from low signal-to-noise ratio (SNR) and limited spectral bands, leading to noisy bathymetry retrievals. To address this challenge, we developed BathyUNet++, a UNet++ architecture enhanced with spatial and channel squeeze & excitation (scSE) attention modules, to jointly leverage the spatial context of neighboring pixels and the complementary spectral information for bathymetry retrieval from SuperDove imagery. The model was trained using matchups of SuperDove patches (32-by-32 pixels) and bathymetry derived from Landsat 8 and airborne LiDAR (ALB). We designed a novel loss function, the Center-Region Mean Square Error (CMSE), which focuses on the center valid receptive field of each patch to effectively reduce instability and checkerboard artifacts often seen in satellite products generated via patch-based processing. Independent validation demonstrates that BathyUNet++ achieves high-accuracy bathymetry retrievals, with median absolute percentage differences of less than 15% when compared against independent depth measurements from ICESat-2 and ALB. Notably, BathyUNet++ substantially reduces noise compared with bathymetry maps derived from pixel-based algorithms, yielding spatially consistent bathymetry maps. The proposed framework overcomes the limitations of low-SNR data and can be readily adapted to other HR sensors, offering a promising solution for the broader application of HR bathymetry retrieval in optically shallow waters.
Wendian Lai*, Xiamen University, [email protected], https://orcid.org/0000-0003-0671-6424
Xiaolong Yu, Xiamen University, [email protected]
Anders Knudby, University of Ottawa, [email protected]
Zhongping Lee, Xiamen University, [email protected]
Poster Session | 1 | 2 | 3 | 4 | Schedule at a Glance
