Ulearn: An explainable uncertainty-aware machine learning tool for unsupervised classification

Abstract

Research and application combining machine and deep learning algorithms and earth observation data has seen tremendous success in the last decades. However, most popular models are black-boxes. A scarcity of eXplainable unsupervised models with adequate accuracy and robustness is still evident. In this research we propose a tool to make unsupervised classification of image data with an eXplainable probabilistic model which makes it feasible for domain experts to validate the results of the classification, especially when the model is not adequately certain of its predictions. The eXplainable model used is a Bayesian generative model called Latent Dirichlet Allocation (LDA). LDA has recently been established as an eXplainable model and able to process diverse data types e.g., text, image. Audio etc. We propose a user interface tool to facilitate using LDA for images classification. We present two use-cases with the tool: 1) automatic coastline change detection with Sentinel-2 images at Baltic Sea, and 2) fire detection at Los Angeles.