Towards Prototype-Based Explainable Machine Learning for Flood Detection

Ivica Obadic

Chair of Data Science in Earth Observation at the Technical University of Munich

The increasingly available high-resolution satellite data has shown to be a valuable resource in tackling pressing issues related to climate change and urbanization such as flood detection. In recent years, deep learning models based on satellite data have shown to be a promising approach for accurate flood detection. Nevertheless, their usage in such critical real-world applications is hindered by the complex and non-linear inference mechanism that results in an intransparent decision process. The goal of this research visit to the Lancaster Intelligent Robotic and Autonomous System Centre is two-fold. First, we aim to shed a light on the workings of the existing state-of-the-art deep learning models on flood detection. Second, we will develop a novel and explainable by-design machine learning model that combines prototype-based explainability with graph neural networks.
The baseline understanding of the state-of-the-art models for flood detection will be obtained by applying established explainable machine learning methods like TCAV. The prototype-based approaches are explainable by design as they decompose the model decisions in terms of high-level human-understandable concepts. Graph neural networks are able to capture the geometry of the objects in a scene which is of high relevance for accurate flood detection. Therefore, we believe that designing a novel approach that combines these techniques has a high potential to outperform the predictive accuracy of the existing models whilst enabling improved and inherent transparency of model decisions.

Keywords: deep learning, unsupervised concept discovery, self-supervised learning, flood detection, earth observation
Scientific area: Explainable Machine Learning