| Title | GeoCore: an Efficient and Scalable Framework to Optimize Geospatial Machine Learning |
|---|---|
| Authors | Ognjen GRUJIC, Thomas HOSSLER, Jacob LIPSCOMB, Rachel MORRISON, Connor SMITH |
| Year | 2025 |
| Conference | Stanford Geothermal Workshop |
| Keywords | Machine Learning, Geothermal, Play Fairway, Great Basin, Modeling, Geospatial, Exploration, INGENIOUS, Python, Codebase |
| Abstract | Machine learning is used extensively in many geospatial applications, including geothermal and mineral resource exploration. The scale on which these models are applied can be quite significant, especially when we consider the need for predictions on very fine grids and across large regions. Developing machine learning (ML) models requires many iterations, experimentation with various model parameters, and management of training and validation sets. In geospatial ML it is also important to consider the proper spatial separation between training, test, and validation sets, all of which significantly increase the computational requirements. In this work, we present an open-source library called GeoCore that enables efficient development of geospatial ML models with an H3 grid indexing of a wide range of resolutions. GeoCore enables modelers to work on top of any spatial database (PostgreSQL, Snowflake, BigQuery) with data indexed with H3 grid blocks, automatic feature caching, and MLflow for experiment tracking. GeoCore includes a dynamic registry system for utilizing various machine learning models, spatial cross-validation techniques, and utilities such as statistical fold plots and automatic buffering with user specified validation sets. Internally, we leverage GeoCore with public and proprietary geospatial data to produce large scale geothermal play fairway analysis models. We will demonstrate the entire modeling process with one of our models using public geospatial data from the INnovative Geothermal Exploration through Novel Investigations Of Undiscovered Systems (INGENIOUS) project. |