The convergence of new spatially aware data sources and compute-intensive machine learning algorithms is transforming how Earth is observed, quantified, and interpreted. As one distinct trajectory of planetary computing, GeoAI is both shaping and being shaped by the forces propelling AI in new directions. While the preceding pages have described and demonstrated how established GeoAI operates, this page serves as an outpost for tracking its next-stage developments and its tentative yet inevitable engagement with generative AI.
This research paper describes features and conditions that make GeoAI a particularly challenging endeavor. In GeoAI, reproducibility and replicability can be compromised not only by model uncertainty, but by the spatial heterogeneity of geospatial data.
GeoAI Reproducibility and Replicability: a Computational and Spatial Perspective. April 15, 2024
This article discusses a new chatbot, GeoGPT, designed by the international science program Deep-time Digital Earth. The article describes the bot's lack of transparency, state censorship, and potential copyright infringement.
Geoscience AI in crisis? June 17, 2024
This thesis offers a technical overview of spatial machine learning frameworks, with a focus on expanding existing multi-modal approaches that combine imagery and text through large language models.
Towards GeoAI Foundation Models: A Multimodal Learning Framework with Spatial Knowledge, 2024
This book, from which several sections are replicated here, places the documentation of agroforestry practices in the Alas Mertajati into the context of planetary computing, large AI systems, digital twins and Earth models.
The Planetary, edited by Nils Gilman, presents important contributions on the condition of Planetarity as a new philosophy in the lineage of massive conceptual change. Here, Planetarity is presented as the Third Great Decentering following Copernican heliocentrism and Darwin’s theory of natural selection. Planetarity supplants the figure of the human as the master of all things.
The Planetary. Berggruen Press, 2025
Perhaps the most significant advancement in the search for a general-purpose Earth model is AlphaEarth Foundations, developed by DeepMind. AlphaEarth is a 64-dimensional embedding field model that integrates spatial and temporal context from multiple sources, including over 3 billion optical, radar, and text datasets from LandSat, Sentinel-2, Sentinel-1, and Wikipedia. According to the project's authors, AlphaEarth outperforms other models in detecting land cover changes. However, it remains to be seen whether its performance will extend to domains with very sparse data or to minority agricultural practices.