Integrating AI into Research Data Workflows: Leveraging ONNX Models within Geo Engine for Analyzing Biodiversity Data

Efficient integration and analysis methods, such as those from artificial intelligence (AI) and machine learning (ML), are essential for uncovering complex patterns and gaining new insights from data in the National Research Data Infrastructure (NFDI). However, incorporating sophisticated AI/ML models into reproducible, scalable, and FAIR (Findable, Accessible, Interoperable, Reusable) research workflows remains a critical hurdle. This contribution presents a novel AI extension to Geo Engine, a cloud-ready platform for processing spatio-temporal data, to support a seamless integration of ML models using the Open Neural Network Exchange (ONNX) format. ONNX provides an open standard for representing machine learning models, enabling interoperability between different ML frameworks and tools. By supporting ONNX, Geo Engine enables researchers to directly integrate pre-trained or custom-built ML models into complex data processing and analysis workflows without being locked into specific ML ecosystems. Thus, it substantially improves the FAIRness of their ML-based workflows. We demonstrate the practical application of this capability in the context of NFDI4Biodiversity, particularly enhancing tools such as the Visualization, Analysis, and Transformation (VAT) tool that runs on top of the Geo Engine platform. The VAT tool aims to connect diverse biodiversity data sources accessible through NFDI infrastructures, providing researchers with interactive exploration and analysis tools. The novel integration of ONNX models through Geo Engine workflows greatly expands VAT's capabilities. For example, researchers can define Geo Engine workflows that retrieve Sentinel-2 satellite imagery, apply an advanced ONNX-based cloud detection model, and process cloud-free, temporally aggregated vegetation indices such as the NDVI. These ML-derived insights can then be combined with species occurrence data from other NFDI4Biodiversity services, all within a unified analytical environment. Geo Engine's ONNX operator enables users to apply machine learning models as part of a workflow. This allows users to seamlessly ingest pre-processed data from different sources, using the existing toolbox. Because of its stream-based processing, Geo Engine can also train models externally on large datasets by feeding data into a training process, for instance, running inside a Jupyter notebook, chunk by chunk. The resulting model, exported and uploaded in the ONNX format, works as a regular operator in a workflow and is thus highly reusable. It can be shared with other users and interchanged with other ONNX models, while the original workflow remains unchanged. In summary, it allows users to use available ML models and create new models in a FAIR way by providing data access, harmonization techniques, and seamless connectivity with ONNX. This extension offers several advantages aligned with NFDI goals: 1. Accessibility: Simplifying the operational use of complex ML models for domain experts by integrating them into a familiar data analysis platform within the NFDI ecosystem. 2. Interoperability & Standardization: Leveraging the ONNX standard, promoting the exchange and reuse of ML models across different research groups and tools. 3. Reproducibility & FAIRness: Encapsulating ML model execution within defined Geo Engine workflows enhances the transparency, reproducibility, and FAIRness of complex analyses, as both the data lineage and the specific model version used are tracked.