River Bathymetry Retrieval From Landsat-9 Images Based on Neural Networks and Comparison to SuperDove and Sentinel-2

Abstract

The Landsat mission has kept an eye on our planet, including water bodies, for 50 years. With the launch of Landsat-9 and its onboard Operational Land Imager 2 (OLI-2) in September 2021, more subtle variations in brightness (14-bit dynamic range) can be captured than previous sensors in the Landsat series (e.g., 12-bit Landsat-8). The enhanced radiometric resolution of OLI-2 appeals to the aquatic remote sensing community because the instrument might be capable of resolving smaller differences in water-leaving radiance. This study evaluates the potential to map river bathymetry from Landsat-9 imagery. We employ a neural network (NN)-based regression model for bathymetry retrieval and compare the results with optimal band ratio analysis (OBRA). The effect of Landsat-9 pan-sharpening on depth retrieval is also examined. In addition, we perform an intersensor comparison with Sentinel-2 and newly available 8-band SuperDoves from the PlanetScope constellation. Depth retrieval results from the Colorado and Potomac Rivers imply that Landsat-9 provided more accurate bathymetry across a range of depths up to 20 m, particularly when pan-sharpened. Downsampling the SuperDove data improved bathymetry retrieval due to enhanced signal-to-noise ratio, most notably in deep waters (maximum detectable depth increased from ∼15 to ∼20 m). Similarly, the enhanced spectral resolution of 8-band SuperDoves improved depth retrieval relative to 4-band Doves. The NN-based model outperformed OBRA by incorporating more spectral information.

Keywords

• Bathymetry
• CubeSats
• landsat-9
• machine learning
• neural networks (NNs)
• pan-sharpening