Cloud-based quantification of Spatial Explicit Uncertainty of Remote Sensing-based Benthic Habitat Classification and its utilization in the context of Active Learning

With the latest advances in cloud image processing, scientists and policy makers have found an effective and robust platform to process vast satellite data in order to be able to map the extent, monitor the condition and create effective protection policies for different ecosystems across the globe. Cloud-based techniques though lack information on the spatial explicit uncertainty of the mapping algorithms. In this study, we present a novel approach on the estimation of uncertainty in a benthic habitat classification context. We explore the benefits of such information in the context of better classification results through an ensemble classifier and the visualization of the uncertain areas in an attempt to provide better maps to the policy makers. The study area consists of Komodo and Wakatobi islands in Indonesia while reference and satellite data come from the Allen Coral Atlas(ACA) project sampling and a six-year PlanetScope composite, free of clouds and optical deep waters Our semi-automated algorithm is divided in three sectors. The first one prepares the data in the context of sampling a number of subsets of reference points according to ACA map products and runs the first classification based on the first subset. The second one aims to help the model re-train itself in a data driven way by accepting training points of the remaining subsets that have mediocre to low uncertainty scores. The uncertainty score is calculated based on probabilistic principles and the theory of Information. The last stage consists of three ensemble classifiers with the inputs of the classification of the second sector. The ensemble classifiers produce three different map products based on mode, max likelihood and simple weighted average values, respectively. According to the results, our workflow is able to minimize the noise of reference points, especially when they come from mapping products and non in-situ measurements. Furthermore, accuracy scores following retraining are better than the initial ones which verifies the hypothesis of removing training data with noise in an attempt to reduce the introduced bias in the classification model. Last but not least, the bi-product of classification uncertainty map can be utilized as a tool for better in-situ sampling planning and render a better understanding to policy makers regarding the validity of scientific reports such as change detection, satellite derived bathymetry and blue carbon accounting, among others.