GeoAI: E-Learning Platform for AI based Geodata Analysis

The increasing availability of geological, environmental, and climate dataset has rendered the traditional analytical approaches insufficient for establishing and interpreting its complex patterns. We currently find ourselves in the era of Artificial Intelligence (AI), which offers the geoscience community an opportunity to identify the non-linear relationships within these datasets, thereby improving the predictive accuracy. However, the adoption of such novel methods comes with its own challenges. In this work, we have identified two primary challenges. First, although it has become increasingly easier to run a basic Machine Learning (ML) algorithm, the lack of understanding of its mathematical foundation and architectural principles poses a challenge to its pragmatic application. Despite the widespread availability of online resources, we have observed that an individual generally finds themselves overwhelmed and unable to translate these methods into practice, largely due to the absence of resources that explains the algorithms directly within the geological context. Second, as a consequence of this limitation, the models are frequently trained in an overly-simplified manner, which leads to compromised results. Careful feature selection, and transformation are critical to deriving meaningful information from complex geo-scientific datasets. Given that such datasets often consist of parameters spanning different scales, failure to appropriately scale and pre-process the data prior to model training has been observed to have significantly impacted the performance metrics. 

To address these challenges, the Freiraum project GeoAI is being carried out at Technische Universität Berlin. The objective of this project is to build an e-learning platform that explains the algorithms starting from ML (Supervised and Unsupervised) to Deep Learning methods. These algorithms will be implemented using diverse datasets commonly employed in geo-scientific studies, sourced from reliable and well-known open-source repositories. Example applications include time-series analysis of meteorological and ground water datasets for continuous prediction, as well as cloud image classification, sound data, among others. The overarching aim of this work is to demonstrate mathematically sound data pre-processing workflows and to provide guidance on selecting an appropriate model for specific tasks. 

This project seeks to make ML methodologies accessible to individuals interested in applying such techniques efficiently in their work, in a manner that is both comprehensible and mathematically rigorous. The platform also accounts for users who prefer limited engagement with algorithmic theory programming, particularly in Python. To accommodate this, reusable and scalable code implementations will be provided, enabling reproducibility across studies involving similar datasets. Additionally, the project actively incorporates feedback from university-level students who are currently being introduced to these topics as part of their academic curricula.

The anticipated impact of the GeoAI platform is informed by the success of a related initiative, SOGA (Statistics and Geodata Analysis using R and Python), developed at Frei Universität Berlin for geo-statistics education. In strong collaboration with the Freiraum project GEOSTAT (FU Berlin), GeoAI aims to provide a holistic perspective on ML models tailored specifically to the geosciences. The platform is planned for public release by January 2027.