Accuracy assessment of annual land cover time series derived from change-based updating

Abstract

The development of temporally consistent land cover time series from satellite-based earth observation has proven difficult due to variability in sensor observations. This leads to spurious land cover differences between maps when standard supervised classification approaches are applied. To reduce this effect, a common solution has been to first detect change and update a base map for only these change areas. Assessing the accuracy of land cover time series is challenging because multiple maps need to be assessed for both land cover classification and change detection accuracies. Regarding a change based updating approach; accuracy is close to that of the original base map for a time series where only a small percent change occurs. Over longer periods where significant change has accumulated the accuracy becomes more dependent on the change and update labeling accuracy. Thus, accuracy for a change based approach can be seen as a function of the base map, change detection, and update accuracies. A specific formulization is developed to summarize these components and applied to investigate accuracy of a 250 m resolution time series for Canada. Results show that the time series accuracy was in a large degree predetermined by the base map accuracy because there was only a small amount of change over the period and the base map and update accuracies were similar. Increasing the update accuracy by a few percent, within the precision of its estimation, would improve the accuracy of the time series evaluated as it is extend in time.