Improving Robustness of Satellite Image Processing Using Principal Component Analysis for Explainability

Kurzfassung

Finding test-cases that cause mission-critical behavior is crucial to increase the robustness of satellite on-board image processing. Using genetic algorithms, we are able to automatically search for test cases that provoke such mission-critical behavior in a large input domain. However, since genetic algorithms generate new test cases using random mutations and crossovers in each generation, they do not provide an explanation why certain test cases are chosen. In this paper, we present an approach to increase the explainability of genetic test generation algorithms using principal component analysis together with visualizations of its results. The analysis gives deep insights into both the system under test and the test generation. With that, the robustness can be significantly increased because we 1) better understand the system under test as well as the selection of certain test cases and 2) can compare the generated explanations with the expectations of domain experts to identify cases with unexpected behavior to identify errors in the implementation. We demonstrate the applicability of our approach with a satellite on-board image processing application.