Shared raster support reveals pseudo-replication in settlement- centered remote-sensing monitoring

Abstract Remote-sensing studies often treat buffers around nearby settlements as separate observations even when they reuse many of the same raster cells. This shared spatial support can create pseudo-replication that is distinct from ordinary geographic autocorrelation. We evaluated this problem for five 2000–2022 landscape-change metrics around 210 historically documented rural settlements in China’s Yellow River Delta. Among 21,945 pairs of 3-km buffers, 2,036 overlapped, and outcome-independent 4-km thinning reduced the cohort to 43–50 sites. We combined overlap auditing with 15-km blocked spatial cross-validation, coordinate-only baselines, mathematical-coupling tests, residual Moran’s I, and low-overlap thinning. Fragmentation was the only metric with positive out-of-fold R² across both predictor branches and both tree algorithms (best R² = 0.388; Spearman ρ = 0.761), while a quadratic coordinate model alone achieved R² = 0.340. Fragmentation remained positive across all six thinning orders, but every full-cohort model retained significant residual spatial autocorrelation. Built-up change, cropland loss, spectral stress, and the composite stress metric were weak, negative, or coupling-sensitive. Shared raster support therefore changes how settlement-centered monitoring should be interpreted: nominal place counts can overstate spatial replication, and predictive screening requires evidence beyond direct descriptive change.