An unsupervised approach for subpixelic land-cover change detection

In this paper, we present a new method for subpixetic land-cover change detection using coarse resolution time series, as they offer a high time-repetitiveness of acquisition. Changes are detected by analyzing the coherence between a coarse resolution time series and a high resolution classification as a description of the land-cover state at the date of reference. To that aim, an a-contrario model is derived, leading to the definition of a probabilistic coherence criterion free of parameter and free of any a priori information. This measure is the core of a stochastic algorithm that selects automatically the image subdomain representing the most likely changes. Some particular problems related to the use of time series are discussed, such as the potential high variability of a time series or the problem of missing data. Some experiments are then presented on pseudo-actual data, showing a good performance for change detection and a high robustness to the considered resolution ratio (between the high resolution classification and the coarse resolution time series).