Simulation Analysis of Error Propagation on Land Cover Change Maps: A Comparison of Contextual and Non-contextual Models

The use of land cover change map is subject to error propagation from multi-temporal land cover classification maps. Understanding the factors determining error propagation to land-cover change maps helps to select appropriate classification models and characterize the associated uncertainties. In this paper, we presented a simulation analysis on the rates of error propagation for both non-contextual and contextual classification models. The simulation approach was based on simulated annealing with careful experimental designs to control two related factors including the spatial and temporal patterns on the errors in spectral probability estimation. The results showed that the two factors had different influences on the error propagation for non-contextual and contextual classification models. For non-contextual models, increasing temporal dependence of errors could reduce the rate of error propagation while spatial dependence of errors did not have an impact on the error propagation. For contextual classification models, the use of spatialtemporal information significantly reduced the rate of error propagation. However, the utilities of the spatialtemporal information in mitigating error propagation were dependent on the spatial dependence of errors. The impact of the temporal dependence of errors was weakened in the contextual models.