First Steps in Estimating the Spatial Uncertainty of Maximum Likelihood Tasks in a Cloud-based Environment in Context of Marine Remote Sensing

Kurzfassung

Recent developments in remote sensing technology including cloud computing and globally available optical satellite archives have allowed access to vast volumes of data, computation and scalability for mapping seagrasses and their environment. , Yet, beyond the traditional accuracy assessment, there is a broader lack of knowledge and methods for the per-pixel uncertainty of remotely sensed seagrass data.Spatially-explicit uncertainty is not only essential for more accurate remote sensing of seagrass extent, health and bathymetry, but could also aid more effective quantification of seagrasses’ ecosystem services like blue carbon stocks and coastal biodiversity maintenance. In this study, we utilise the open satellite image archives of Sentinel-2 and PlanetScope, through the Google Earth Engine (GEE) platform to develop per-pixel uncertainties of thematic benthic habitat mapping and continuous satellite based bathymetry data according to machine learning probabilistic principles.We present our uncertainty metrics and applications in two nationwide case studies in Bahamas and Belize. In contrast to traditional approaches that estimate uncertainty for the whole image/distribution, our approach, quantifies the uncertainty per pixel of both thematic and continuous remotely sensed data across large spatial scales and up to 5 m resolution. Our approach can improve the confidence and scalability of large-scale assessments of seagrass extent, condition and ecosystem services, supporting more effective policy uptake of seagrass ecosystems.