Abstract: An adaptive sampling strategy is proposed for location-specific characterization of accuracy in land cover change information. The local accuracy characterization strategy was established based on local patterns of land cover change maps (e.g land cover change classes, patch size, heterogeneity and dominance), which include exploring covariates significantly relate to accuracy. Standard error of prediction accuracy was used for identifying the area which needs to improve the reliability of prediction accuracy and locating samples adaptively and progressively. The performance of different sampling methods for accuracy prediction was evaluated at the same testing samples in Wuhan. It was indicated that 100 more training samples selected by adaptive sampling strategy lead to about 50.66% increase in prediction accuracy, as measured by sums-of-squares. In comparison, for random sampling, the same increase in training sample size led to about 17.22% increase in prediction accuracy, as measured by sums-of-squares. This confirms that adaptive sampling strategy improves the sampling efficiency while reduces the uncertainty in local accuracies prediction. Model selection reveals that land cover change classes and dominance are the highest significant covariates.

Key words: land cover change, accuracy, logistic regression, sampling