Spatially-Explicit Uncertainty of Remote Sensing Coastal Biodiversity Products using a scalable cloud-based framework in the Google Earth Engine

Kurzfassung

Recent advances in remote sensing have enabled the global monitoring of Earth's biodiversity. These developments are providing global information on the extent, structure, function and services of different ecosystem types, and their benefits to the environment and humans. In contradiction with the advances, relevant uncertainty methods and information are missing the understanding of the product biases. In our study, we present a uncertainty quantification framework, developed entirely within the Google Earth Engine, which assesses both thematic (e.g., ecosystem presence/absence) and continuous products (e.g., satellite-derived bathymetry) related to coastal biodiversity using multi-temporal and cloud-free 10-m Sentinel-2, field data collections, and human-annotated data points. By exploiting the cloud-native machine learning classifier and its outputs, we estimate the uncertainty of the procedure per pixel. With that information, our model is able to re-train itself in a data driven way and produce better results. There are three areas of interest in this study. The first is the Archipelago of Bahamas, where we assess a four-class benthic habitat classification product. Our second and third study area is the national scale of Belize and the Quirimbas Archipelago (Mozambique), respectively, in which we generate a satellite-derived bathymetry map. In the case of classification, our model achieved a better overall accuracy in comparison with the initial classification while the producer and user accuracy of the habitat class that we are interested in, seagrass, rose by 13% and 7% respectively. On the regression results, our framework highlights the areas with most uncertainty given the byproducts of the maximum likelihood regression that took place. While still in its alpha version, we think that further developments of the framework could allow better quantification of the data and model uncertainty. By reducing the uncertainties in the coastal biodiversity monitoring, more effective policy making efforts can be achieved and thus, better conservation.