Predicting out-of-domain performance under geographic distribution shifts (Papers Track)
Haoran Zhang (Harvard University); Konstantin Klemmer (Microsoft Research); Esther Rolf (University of Colorado, Boulder); David Alvarez-Melis (Harvard University)
Abstract
In machine learning for geographic data, we often observe differences in data availability and distribution shifts across distinct geographic units, e.g., continents. This is a common challenge in remote sensing tasks, such as crop yield forecasting or flood mapping. In many of these scenarios, we have models trained on a data-rich region and apply domain adaptation to transfer predictive capabilities to the target region. However, the effectiveness of domain transfer can suffer from distribution shifts, posing critical challenges for model deployment. In this work, we show that, even in the absence of labels, certain domain distance measures, based on image and location embeddings, can serve as a proxy measure for transfer performance. We further highlight this capacity on a set of real-world geographic adaptation datasets, spatial splits for domains, and models for adaptation training.