Bathymetry retrieval from CubeSat image sequences with short time lags

Abstract

The rapid expansion of CubeSat constellations could revolutionize the way inland and nearshore coastal waters are monitored from space. This potential stems from the ability of CubeSats to provide daily imagery with global coverage at meter-scale spatial resolution. In this study, we explore the unique opportunity to improve the retrieval of bathymetry offered by CubeSats, specifically those of the PlanetScope constellation. The orbital design of the PlanetScope constellation enables the acquisition of image sequences with short time lags (from seconds to hours). This characteristic allows multiple images to be captured during a short period of steady bathymetric conditions, especially in dynamic environments like rivers. We hypothesize that taking the ensemble mean of a CubeSat image sequence can enhance bathymetry retrieval compared to standard single-image analysis. Along with the existing optimal band ratio analysis (OBRA) algorithm, we also use a new neural network-based depth retrieval (NNDR) technique to infer bathymetry from both individual and time-averaged images. The two methodologies are evaluated using field data from five different river reaches with depths up to 15 m and both top-of-atmosphere (TOA) radiance and bottom-of-atmosphere (BOA) surface reflectance PlanetScope data products. Despite low spectral resolution and concerns about the radiometric quality of CubeSat imagery, accuracy assessment based on in-situ comparisons indicates the potential (0.52 < R2 < 0.7 for the NNDR method) of PlanetScope imagery to retrieve depths up to ∼ 10 m in clear water conditions. The proposed image averaging consistently improves bathymetry retrieval over single image analysis. The NNDR technique was found to outperform OBRA, illustrating the importance of leveraging all spectral bands through machine learning approaches. TOA data provided more robust bathymetry results than BOA data for the OBRA technique, but the NNDR technique was minimally impacted by the type of data product.