No Location Left Behind: Measuring and Improving the Fairness of Implicit Representations for Earth Data

TL;DR

This paper tackles fairness in implicit neural representations (INRs) for Earth data by introducing FAIR-Earth, a high-resolution, open-source evaluation framework that enables per-subgroup fairness analyses across multiple modalities. It demonstrates that state-of-the-art INR encodings exhibit significant local biases, especially for high-frequency, localized features such as islands and coastlines, even when global performance is strong. To address these disparities, the authors introduce Spherical Wavelet encodings, which provide multi-scale, localized representations on the sphere and empirically show reduced per-group biases while remaining competitive with traditional baselines. The work delivers a practical, open framework for equitable Earth INRs and points to future refinements in multi-scale spherical encodings to further improve fairness and robustness in geospatial modeling.

Abstract

Implicit neural representations (INRs) exhibit growing promise in addressing Earth representation challenges, ranging from emissions monitoring to climate modeling. However, existing methods disproportionately prioritize global average performance, whereas practitioners require fine-grained insights to understand biases and variations in these models. To bridge this gap, we introduce FAIR-Earth: a first-of-its-kind dataset explicitly crafted to examine and challenge inequities in Earth representations. FAIR-Earth comprises various high-resolution Earth signals and uniquely aggregates extensive metadata along stratifications like landmass size and population density to assess the fairness of models. Evaluating state-of-the-art INRs across the various modalities of FAIR-Earth, we uncover striking performance disparities. Certain subgroups, especially those associated with high-frequency signals (e.g., islands, coastlines), are consistently poorly modeled by existing methods. In response, we propose spherical wavelet encodings, building on previous spatial encoding research. Leveraging the multi-resolution capabilities of wavelets, our encodings yield consistent performance over various scales and locations, offering more accurate and robust representations of the biased subgroups. These open-source contributions represent a crucial step towards the equitable assessment and deployment of Earth INRs.

Figures (25)

• Figure 1: Heatmap of the spatial distribution of approximation errors using existing INRs to model land-sea data of the Earth. Leveraging the resolution of FAIR-Earth, we uncover clear bias against islands, where error magnitude is significantly higher. Against the same task, Spherical Wavelets resolves these issues by reconciling global signals with fine and localized signals. Details and plots available in \ref{['APP:Misc Figures']}.
• Figure 2: FAIR-Earth is the first Earth INR-framework for fairness assessment, and moreover is an open-source package that allows for agile testing and analysis of subgroup-level performance.
• Figure 3: Visualization of existing encoding mechanisms.
• Figure 4: Model behavior at different resolutions and regularizations. Smaller models fail to capture fine/local signals. Larger models poorly reconcile local signals with existing global signals.
• Figure 5: Qualitative performance of Spherical Harmonic on surface temperature regression task. SH exhibit bias against high-frequency details, with loss concentrated in areas of abrupt change (e.g., coastlines).