Large Area Mapping of Annual Land Cover Dynamics Using Multitemporal Change Detection and Classification of Landsat Time Series Data

Abstract

. Land cover characteristics remain of particular interest to the monitoring and reporting communities, and approaches for generating annual maps of land cover informed by change information derived from long time series are critically needed. In this study, we demonstrate and verify the utility of disturbance and recovery metrics derived from annual Landsat time series to inform the classification of annual land cover over a > 1.2 million hectare forest management area in the Boreal Mixedwood Region of northern Ontario, Canada. Annual land cover maps were generated, producing temporally informed products and compared to the established approach of using single-date spectral variables and indices. The Random Forest (RF) classification algorithm was used to classify land cover annually between 1990 and 2010, followed by the application of an annual temporal filter to remove illogical land cover transitions. Change detection in the study area had an overall accuracy of 92.47%. The use of time series metrics in the classification of land cover improved overall accuracy by 6.38% compared to single-date results. Using a separate independent reference sample, the RF classification approach combined with postclassification transition filtering resulted in an overall classification accuracy of 87.98%. The use of annual change and trend information to guide land cover, which is further informed by logical land cover transition rules, points to the creation of efficient, robust, and reliable land cover products in a transparent and operational fashion.