Modeling Spatiotemporal Heterogeneity in Earth Science Machine Learning: An End-to-End Approach

Siqi Du, Hongsheng Huang, Kaixin Shen, Shengjun Tang

Keywords: AI4Science, Remote Sensing, Tabular Machine Learning, Spatiotemporal Prediction, Spatiotemporal Heterogeneity

TL;DR: We propose an end-to-end model for spatiotemporal prediction in Earth Science, using a generative framework to capture global patterns and sparse spatial heterogeneity, while mitigating overfitting.

Abstract: In Earth sciences, unobserved factors often lead to spatially nonstationary distributions, causing relationships between features and targets to vary across locations. Traditional tabular machine learning methods struggle to effectively model this spatial heterogeneity. While approaches like Geographically Weighted Regression (GWR) capture local variations, they often miss global patterns, overfit local noise, and lack the ability to model temporal changes in spatial heterogeneity. Our research aims to model spatiotemporal heterogeneity. To achieve this, we propose an end-to-end approach that fits the entire dataset to capture global patterns, while designing the model as a conditional generative framework to learn sparse spatial heterogeneity, mitigating overfitting through localized condition sharing. Our method involves four key steps: constructing a spatiotemporal graph, encoding tabular features, aggregating spatial heterogeneity node embeddings via graph convolutions, and decoding with spatial condition vectors for location-specific predictions. We validate our approach by predicting vegetation gross primary productivity (GPP) using global climate and land cover data (2001–2020). Trained on 50M samples and tested on 2.8M, our model achieves an RMSE of 0.836, outperforming GWR (2.149), LightGBM (1.063) and TabNet (0.944). Visual analysis of the learned node embeddings reveals clear spatial heterogeneity patterns and their temporal dynamics.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 13346