How Many Independent Samples Does a Satellite Image Contain? Generalization Bounds for Spatially Dependent Data
Abstract
Machine learning classifiers for remote sensing imagery are typically evaluated as though every pixel were an independent sample. Spatial autocorrelation violates this assumption, since neighboring pixels carry redundant information which inflates sample sizes. How many independent samples does a satellite image actually contain? For an n × n image whose spatial correlation persists over a range of r pixels, the effective sample size is Θ(n2/r2), not n2. We prove this as a finite-sample upper bound for classifiers on spatially correlated data, and show via a matching lower bound that the rate is tight, and no algorithm can do better. We extend the results to images with directional correlation and spatially varying correlation structure. Our result justifies spatial cross-validation since block holdout with separation proportional to the correlation range achieves optimal generalization guarantees, while random holdout can underestimate confidence interval widths by a factor proportional to r. We validate the theory on synthetic data and satellite image tiles from three sensors (Landsat 8, Sentinel-2, and Sentinel-1).
Versions
Related
- Independent Verification of the European Climate Law's Effect on Nitrogen Dioxide Pollution: A Satellite-Derived, Difference-in-Differences Analysis
- Robust Inference for Dyadic Data with Spatially Dependent Nodes
- Decay Rates and Domain Dependence of a Coupled Wave-Heat System with Spatially Dependent Heat Coefficients
- An Analysis of Self-supervised Pre-training with Dependent Samples
- Rigidity results for overdetermined problems of p-Laplace equation with spatially dependent data
BibTeX
@misc{young2026many,
title = {{How Many Independent Samples Does a Satellite Image Contain? Generalization Bounds for Spatially Dependent Data}},
author = {Robin Young},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.08227},
eprint = {2610.08227},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.08227},
}
RIS
TY - UNPB
TI - How Many Independent Samples Does a Satellite Image Contain? Generalization Bounds for Spatially Dependent Data
AU - Robin Young
PY - 2026
DA - 2026/10/06
PB - arXiv
DO - 10.48550/arxiv.2610.08227
UR - https://arxiv.org/abs/2610.08227
AB - Machine learning classifiers for remote sensing imagery are typically evaluated as though every pixel were an independent sample. Spatial autocorrelation violates this assumption, since neighboring pixels carry redundant information which inflates sample sizes. How many independent samples does a satellite image actually contain? For an n × n image whose spatial correlation persists over a range of r pixels, the effective sample size is Θ(n^2/r^2), not n^2. We prove this as a finite-sample upper bound for classifiers on spatially correlated data, and show via a matching lower bound that the rate is tight, and no algorithm can do better. We extend the results to images with directional correlation and spatially varying correlation structure. Our result justifies spatial cross-validation since block holdout with separation proportional to the correlation range achieves optimal generalization guarantees, while random holdout can underestimate confidence interval widths by a factor proportional to r. We validate the theory on synthetic data and satellite image tiles from three sensors (Landsat 8, Sentinel-2, and Sentinel-1).
ER -