Explainable Artificial Intelligence in Remote Sensing

Abstract

Although artificial intelligence methods have achieved notable success in many sectors, there is an increasing demand towards explainability and trustworthiness of these methods. Currently, machine learning models such as deep learning are “opaque”, such “opacity” introduced by sub-symbolism. We attempt to explore explainable machine learning methods in understanding Earth observation data. Recently, there have been efforts towards explaining machine learning (X-ML) models, contributing to the paradigm of eXplainable Artificial Intelligence(XAI). The important conceptual propositions X-ML are interpretability, transparency and explainability. Here, we propose an explainable data mining approach based on Latent Dirichlet Allocation (LDA) which supports the idea of transparency, interpretability and explainability. Informally, transparency is the ability to understand the mechanism of each component of a method. Researchers delineate three levels of transparency: design transparency, algorithmic transparency and model transparency. Design transparency is concerned with clear logic behind design decision such as model parameters. Algorithmic transparency is the ability to understand how the algorithm works from a mathematical point of view. A model is called algorithmically transparent if input-out relation and the process can be written down as mathematical formula. Model transparency ensures traceability of the outcomes. We demonstrate of adherence of each step of our 5-step method to design, model and algorithmic transparency. The second aspect of explainable machine learning is interpretability, which is nothing but making sense of intermediate outcomes e.g., latent layers from the model in combination with the help of domain knowledge. We use a exploit the latent variables retrieved from the LDA to create a interpretable visualizations for non-visual Sentinel-1 data and demonstrate the idea of interpretability with the help of product quick-look and classification maps from another research. We define three levels of interpretability: low, medium and high. Low interpretability only uses the visualization, medium interpretability compares the visualization with the quick-look and high interpretability is achieved by comparing the visualization with the product quick-look and the classification map. The third aspect, explainability refers to the decision made by using features in the interpretable domain. We deliver explainable class similarity measures and semantic relations based on seamless latent variables retrieved from the LDA model. In general, the method proposes a LDA-based data mining approach as a contribution towards XAI in the remote sensing field.