The effects of different classification models on error propagation in land cover change detection

Abstract

The use of land cover change maps is subject to the propagation of errors involved in classifying multi-temporal land cover maps. Understanding the link between classification processes and error propagation helps to determine appropriate classification models to mitigate the error propagation rate. In this paper, we present a simulation analysis on error propagation in land cover change detection using three classification models: a non-contextual model, a contextual model based on spatial smoothing, and a contextual model based on Markov random fields (MRF). A spatial simulation approach based on simulated annealing was developed with careful experimental designs to control two related factors including the spatial/temporal patterns of estimation errors associated with spectral probabilities. The results showed that the contextual classification model based on MRF had the smallest error propagation rate while the non-contextual classification model had the largest rate under all scenarios. The two factors had different effects on the error propagation for different classification models. For the non-contextual model, increasing temporal correlation of errors could reduce the error propagation rate while spatial autocorrelation of errors did not have a big impact on the error propagation. For the two contextual classification models, the use of contextual information significantly reduced the error propagation rate. However, the value of contextual information in mitigating error propagation was highly dependent on the spatial autocorrelation of the errors. The impact of the temporal correlation of errors was weakened in the contextual models.