A satellite image holds far fewer independent samples than pixels: correlation range sets the count, and random holdout overstates accuracy
Synopsis
Remote-sensing classifiers are usually evaluated as if every pixel were an independent sample, but neighboring pixels are strongly correlated, so that assumption fails. A finite-sample upper bound with a matching lower bound shows an image with correlation range r carries about n/r² effective samples rather than n, and this rate is minimax tight. Block holdout with separation proportional to the correlation range then attains optimal generalization guarantees, while random holdout understates confidence-interval widths by roughly a factor of r.
Interpretation
The independent information in a satellite image is set by the spatial correlation range, not by the pixel count. Earlier learning-theoretic guarantees assume independent and identically distributed samples, and remote-sensing assessments often compute error bars from the nominal pixel count, so neither applies under spatial autocorrelation. On synthetic data and satellite image tiles from three sensors, Landsat 8, Sentinel-2 and Sentinel-1, the measured effective sample size falls as a power law in block size with slopes matching the predicted rate.
The approach cuts the image into small blocks separated by gaps so the blocks are nearly independent, then applies standard concentration inequalities for independent samples. Larger blocks reduce variance but inflate the pseudo-dimension of the block-averaged loss class and reduce the block count, while wider gaps improve independence but consume pixels. Optimizing this tradeoff sets the optimal gap at the correlation-range scale and reduces the image to about n/r² nearly independent blocks; for directional correlation and piecewise-stationary scenes the blocks become rectangular or the gaps are set per stratum.
Perspective
The results apply to strictly stationary random fields satisfying the spatial mixing condition, with the correlation range estimated from an empirical variogram; anisotropic fields use rectangular blocks and direction-dependent gaps, and piecewise-stationary scenes combine per-stratum ranges through a variance-weighted formula. The intended readers are researchers and engineers working on remote-sensing classification and accuracy assessment, who can use the prescription to set spatial cross-validation fold separation and correct confidence intervals.
The upper bound is worst-case and conservative, with unoptimized constants and polylogarithmic factors, so the value is the scaling rather than a sharp constant. The correlation range must be estimated from an empirical variogram, and a poorly specified variogram biases the effective sample size; the bounds are driven by the mixing rate of the loss field, whereas the measured variograms use raw features and may overestimate the loss-field range. Label noise is assumed conditionally independent given features, and tasks with spatial label structure may exhibit different mixing behavior.
