Color Me Optically Shallow: A Simple And Adaptive Method For Standardized Analysis Ready Data For Coastal Ecosystem Assessments

Kurzfassung

Coastal aquatic remote sensing (RS) can help monitor the immensely valuable ecosystems of the global seascape, such as seagrasses and corals, by providing information on their extent, condition (e.g., water quality, bathymetry), ecosystem services (e.g., carbon sequestration, biodiversity maintenance), and trajectories. Unlike terrestrial RS, coastal aquatic RS applications require an additional consideration of the water column and its interactions with the light signal. This introduces new challenges as the water column attenuates light differently across the wavelengths, which has implications for signals from the benthic seabed where these subtidal ecosystems thrive. When the object(s) of interest is located on the benthic floor and not floating near the water surface, the additional depth increases the influence of the water column on light and affects the signals sensed by satellites at the top of the atmosphere. Besides these, other effects such as turbidity, waves, and sunglint introduce wide-ranging reflectance values as well. While these challenges have been traditionally handled through often complex methods in local computing environments, contemporary advances in cloud computing and big satellite data analytics offer highly scalable and effective solutions within the same context. The parallel processing of cloud platforms like the Google Earth Engine allows multitemporal composition of thousands of satellite images in a defined area over a defined time range through highly efficient statistical aggregations. As such, this approach yields Analysis Ready Data which are less redundant and more time efficient than the conventional laborious manual search for suitable single satellite image(s) which is often a yearlong assessment over cloud-dense coastal regions like the tropics. Regardless of the method, the pre-processing of the image and/or image composite remains a critical component of a successful coastal ecosystem assessment using RS. The impact of light attenuation changes the returning spectral signal, resulting in different signal profiles for the same seabed cover at different depths. In particular, at deeper depths, darker covers such as vegetated coastal beds (e.g., dense seagrass, microalgal mats) and optically deep water pixels are more likely to be confused and misclassified. A possible solution is to identify and remove these deep water pixels, where the water is too deep and thus no bottom signals are able to return to the sensor. By using a HSV-transformed B1-B2-B3 false-colour composite, namely the hue and saturation bands, of the Sentinel-2 image archive within the cloud computing platform of the Google Earth Engine, we are able to disentangle optically deep from optically shallow waters across four sites (Tanzania, the Bahamas, Caspian Sea (Kazakhstan) and Wadden Sea (Denmark and Germany)) with wide-ranging water qualities to improve the optically shallow benthic habitat classification. Furthermore, we compare our method with the three band ratios from a combination of the same three bands. While the band ratios may perform better in some sites, the specific band combination is site specific and thus might perform worse in others. In comparison, the hue and saturation bands show more consistent performance across all four sites. By using simple statistical reduction, the multitemporal composite is able to automatically mitigate common coastal aquatic RS showstoppers like clouds, cloud shadows or other temporal phenomena. However, there is also a need to remove images with explicitly no useful information, so that it does not affect the statistical approach. The use of metadata properties in the image archive is therefore additionally needed to filter out “bad” images, reducing the unnecessary computational costs of processing these low quality images. Case in point, this is a recommended procedure to filter for lower cloud covers prior to multitemporal composition in Google Earth Engine. We extend this approach further by integrating the various solar and viewing angles to estimate the presence of sunglint, on the basis that the spectral reflectance angle of the scene is a major factor to sunglint presence in satellite images. Finally, we draw comparisons with less pre-processed composites, showcasing methodological benefits for national coastal ecosystem assessments in the Bahamas, Seychelles, and East Africa.