A method for estimation of a land-cover change matrix from error-prone unit-level observations

Abstract

Coregistration and classification errors can seriously compromise direct unit-level (pixel) estimation of land-cover change from remotely sensed data. A more robust alternative to a pixel-based estimation of change is warranted. In a proposed method, spatially adjacent pixels are grouped into 3 × 3 clusters, and the change matrix is obtained from cluster-specific and land cover specific pixel counts at two points in time. The diagonal of a change matrix is estimated by combining an estimate of the temporal correlation of cover type specific, cluster-level counts with an estimate of the odds ratio of no change. Off-diagonal elements are least-squares solutions to a set of linear constraints or obtained by iterative proportional fitting under a model of quasi-independence. In a study with data from five sites, the proposed method produced less biased estimates on three sites if the mean coregistration error was in excess of 0.3–0.7 pixels and on four sites if classification accuracy dropped below 0.9.