A cloud-based framework for the quantification of the spatially-explicit uncertainty of remotely sensed benthic habitats

• Cloud-based framework for the quantification of uncertainty based on the confidence of the classifier. • Seagrass accuracy improves by 1.16–4.77 % (user's) and 4.36–8.54 % (producer's) over a standard classification with RF. • PUNC maps can guide fieldwork and marine planning by showing classification confidence levels. • Compatible with Sentinel-2 and PlanetScope, enabling scalable, time-series habitat analysis. The significant advances of cloud-based remote sensing frameworks have allowed researchers to develop large-scale analytics for better understanding, monitoring of, and decision-making around sensitive and valuable coastal ecosystems like seagrass meadows. However, an information gap related with the spatially-explicit accuracy of Machine Learning (ML) products has been identified. The goal of this study is to estimate the per pixel uncertainty of a Random Forest classification of four benthic habitats and exploit it to retrain the model through training data selection by bootstrapping and producing an ensemble model. The calculation of the spatially-explicit uncertainty is based on the Shannon Entropy equation and the probability values of a successful prediction according to the ML model. The remote sensing data for this study are sourced from the European Union Copernicus Sentinel-2 twin satellite system and Planet’s cubesat satellite constellation respectively, and have been processed and analyzed through the Google Earth Engine cloud-based platform. The national extent of The Bahamas and the regional extent of the Wakatobi archipelago in Indonesia comprise our study sites. Our results indicate the potential of the presented uncertainty workflow for optimizing the classification and the usefulness of the produced uncertainty map to aid policy-makers through our provided spatially-explicit accuracy metrics. More precisely in the case of the Bahamas, the percentile differences for seagrass user and producer accuracies are improved in the ranges of 1.16–4.77 % and 4.36–8.54 %, respectively, in comparison with a standard supervised classification. In conclusion, spatially-explicit uncertainty information can and should be used as unique and vital geospatial information suitable for ML classification optimization and as a tool for better decision-making and field expedition planning, and understanding of benthic ecosystems.