Improving on the explanation of patterns learned by artificial neural networks trained to detect atmospheric features

Reliable detection of atmospheric features such as tropical cyclones, atmospheric rivers, and atmospheric surface fronts is important for weather forecasting and climate analysis. Traditionally, this detection relies on expert knowledge or rule‑based approaches. Artificial neural networks (ANNs) offer a new method for detection. While they detect atmospheric features fast and often accurate, they are black‑box systems, making it difficult to assess whether they identify atmospheric features based on physically meaningful patterns or spurious correlations in the data. Such physically implausible patterns can lead to reduced performance on unseen data even though the ANN performs well in a test setup. Explainable artificial intelligence methods aim to open this black box. Previous studies have adapted Layer‑wise Relevance Propagation (LRP), one of these methods, for explaining ANNs trained for atmospheric feature‑detection. While these studies showed that the extraction of learned patterns is feasible, they also demonstrated that interpreting especially the structure of large‑scale detection patterns remains challenging. To address this, we adapt Concept Relevance Propagation (CRP) for the detection of atmospheric features. Unlike LRP, CRP decomposes the ANN’s detection into concepts, each representing a pattern used by the ANN. Using CRP, we analyze an ANN trained to detect tropical cyclones and atmospheric rivers examining both patterns found detecting singular features and patterns found across the dataset. As an example, for atmospheric rivers, CRP reveals that, in addition to known patterns such as elongated moisture bands and strong winds, the ANN also considers surrounding dry‑air regions - an aspect not previously found. This adaptation of CRP improves the interpretability of ANNs for atmospheric feature detection, enabling a clearer assessment of whether the networks rely on physically plausible patterns or on implausible correlations in the data.