Day 4

Detailed paper information

Back to list

  1. Spyridon Christofilakos DLR - German Aerospace Center Speaker
  2. Alina Blume German Aerospace Center (DLR)
  3. Chengfa Benjamin Lee DLR - Deutsches Zentrum für Luft- und Raumfahrt e.V.
  4. Avi Putri Pertiwi German Aerospace Center
  5. Dimosthenis Traganos DLR - Deutsches Zentrum für Luft- und Raumfahrt
  6. Peter Reinartz DLR - German Aerospace Center
Form of presentation Poster
  • C1. AI and Data Analytics
    • C1.07 ML4Earth: Machine Learning for Earth Sciences
Abstract text The necessity of monitoring and expanding the existing Marine Protected Areas has led to vast and high-resolution map products which, even if they feature high accuracy, they lack information on the spatially explicit uncertainty of the habitat maps, a structural element in the agendas of policy makers and conservation managers for designation and field efforts.The target of this study is to fill the gaps in the visualization and quantification of the uncertainty of benthic habitat mapping by producing an end-to-end continuous layer using relevant training datasets.
To be more accurate, by applying a semi-automated function in the Google Earth Engine’s cloud environment we were able to estimate the spatially explicit uncertainty of a supervised benthic habitat classification product. In this study we explore and map the aleatoric uncertainty of multi-temporal data driven, per-pixel classification in four different case studies in Mozambique, Madagascar, Bahamas, and Greece, which are regions known for their immense coastal ecological value. Aleatoric uncertainty, also known as data uncertainty, is part of the information theory that seeks for the data driven random and inevitable noise under the spectrum of bayesian statistics.
We use the Sentinel 2 (S2) archive in order to investigate the adjustability and scalability of our uncertainty processor in the four aforementioned case studies. Specifically, we use biennial time series of S2 satellite images for each region of interest to produce a single, multi-band composite free of atmospheric and water column related influences. Our methodology revolves around the classification process of the mentioned composite. By calculating the marginal and conditional distribution’s divisions given the available training data, we can estimate the Expected Entropy, Mutual Information and Spatially Explicit Uncertainty of a maximum likelihood model outcome.

Expected Conditional Entropy
Predicts the overall data uncertainty of the distribution P(x,y), with x:training dataset and y:model outcome.
Mutual Information
Estimates in total and per classified class the level of independence and therefore the relation of y and x distributions.
Spatially Explicit Uncertainty
A per pixel estimation of the uncertainty of the classification.

The aim by implementing the presented workflow is to quantitatively identify and minimize the spatial residuals in large-scale coastal ecosystem accounting. Our results indicate regions and classes with high and low uncertainty that can either be used for a better selection of the training dataset or to identify, in an automated fashion, areas and habitats that are expected to feature misclassifications not highlighted by existing qualitative accuracy assessments. By doing so,we can streamline more confident, cost-effective, and targeted benthic habitat accounting and ecosystem service conservation monitoring , resulting in strengthened research and policies, globally.