From Machine Learning to Large-Scale EO Products: Best Practices for Making Maps

Producing large-scale geospatial maps from Earth observation (EO) data has become increasingly accessible due to advances in machine learning (ML) and cloud computing. However, the lack of standardized best practices often leads to errors that propagate through the pipeline, from initial data selection to final map validation. This paper outlines a systematic approach to these challenges, focusing on the interdependencies between six key stages: data infrastructure, preprocessing, dataset construction, uncertainty quantification, inference, and validation.

• Data Infrastructure and Preprocessing: Choices regarding data providers and preprocessing steps are not neutral; they encode assumptions that the model inherits. Practitioners must account for sensor-specific artifacts and ensure that the preprocessing chain used during training is computationally feasible to replicate at a global scale during inference.
• ML Dataset Construction: Spatial autocorrelation is a critical challenge in EO data. Standard random splits often lead to inflated performance metrics; instead, practitioners should use spatial block cross-validation to ensure the model generalizes to new, unseen regions.
• Uncertainty and Validation: A map without quantified uncertainty is of limited scientific value. The authors highlight that uncertainty quantification must account for spatial autocorrelation, as failing to do so leads to significantly underestimated confidence intervals. Furthermore, independent map validation using probability-based sampling is essential to distinguish between model performance on training data and the actual accuracy of the final map product.

Global-scale inference is primarily an engineering challenge governed by data gravity. Whether to move data to compute or compute to data is a fundamental decision that dictates the feasibility of the project. Additionally, practitioners must implement robust artifact mitigation strategies—such as overlapping patches or center-cropping—to handle the spatial biases that often emerge when merging tiled model outputs into a seamless global map.