An object-based change detection method accounting for temporal dependences in time series with medium to coarse spatial resolution

Tracking land cover changes using remotely-sensed data contributes to evaluating to what extent human activities impact the environment. Recent studies have pointed out some limitations of single-date comparisons between years and have emphasized the usefulness of time series. However, less effort has hitherto been dedicated to properly account for the temporal dependences typifying the successive images of a time series. An automated change detection method based on a per-object approach and on a probabilistic procedure is proposed here to better cope with this issue. This innovative procedure is applied to a tropical forest environment using high temporal resolution SPOT-VEGETATION time series from 2001 and 2004 in the Brazilian state of Rondonia. The principle of the method is to identify the objects that most deviate from an unchanged reference defined by objective rules. A probabilistic changed-unchanged threshold provides a change map where each object is associated with a likelihood of having changed. This improvement on a binary diagnostic makes the method relevant to meet the requirements of different users, ranging from a comprehensive detection of changes to a detection of the most dramatic changes. According to the threshold value, overall accuracy indices of up to 91 % were obtained, with errors involving change omissions for the most part. The isolation of changes within objects was made possible through a segmentation procedure implemented in a temporal context. In addition, the method was formulated so as to differentiate between inter- and intra-annual vegetation dynamics. These technical peculiarities will likely make this analytical framework suitable for detecting changes in environments subject to a strongly marked phenology. (C) 2008 Elsevier Inc. All rights reserved.