Landt-Hayen, Marco (2023) Exploring Methods of Explainable AI - Data-driven Attribution of Climate Events. (PhD/ Doctoral thesis), Christian-Albrechts-Universität zu Kiel, Kiel, Germany, 103 pp. . Kiel Computer Science Series, 2023/5 .

Preview

Text Dissertation_MarcoLandtHayen_final.pdf - Published Version Available under License Creative Commons: Attribution-Noncommercial-No Derivative Works 4.0. Download (15MB) | Preview

Abstract

In this work, we explore methods of explainable AI (xAI) to better understand how artificial neural networks (ANNs) come to their conclusion and to visualize the relationship between input and output. Moreover, we aim to use our insights to improve our understanding of the climate system. Working with observational data in the context of geophysics can be challenging since we have consistent data only from the recent past. But most machine learning (ML) methods are data hungry and need sufficient training data. In practice, there exist some exceptions. Among these, we find Echo State Networks (ESNs) as a certain type of recurrent neural networks. We show that layer-wise relevance propagation (LRP) can be used for ESNs. However, there are pitfalls in terms of model-inherent artifacts included in the obtained relevance maps. We revise LRP on various ANN architectures and give guidance on how to control model focus. Furthermore, we address the problem of insufficient training data and introduce a new benchmark data set. In particular, we create a collection of climate indices used to describe the dynamics of the Earth system. In order to have consistent data over a sufficiently long time span, we favor to work with data from modern Earth System Models. The new data set opens the door to tackle another problem regarding geospatial data. In the context of geophysics, we often have missing values. We use U-Net models from the domain of computer vision to reconstruct missing data in two-dimensional geospatial fields. Moreover, we introduce a technique to identify grid points that are most relevant for successful reconstruction. This helps to answer the question where to place a limited number of survey stations to get most information out of it. Eventually, we extend our approach to predict future geospatial fields from sparse inputs. We try to infer climate events like e.g., El Nino or rainfall in the African Sahel region from reconstructed data.

Actions (login required)

View Item