Combining Medium and High Resolution Data in a Multi-scale Approach to Detect Breaks in Satellite Image Time Series
Dutrieux, Loc; Verbesselt, Jan; Kooistra, Lammert; Herold, Martin
GRS Laboratory, Wageningen University, NETHERLANDS

Monitoring of forest dynamics has gained in importance in recent years as change in forest cover is likely to affect aspects of the biosphere such as carbon cycle and biodiversity. Remote sensing time series are very well suited to monitor land dynamics such as deforestation thanks to their systematic and consistent acquisition of data and their archives dating back to the 80's. However, there is usually a challenge that consists in extracting the right information from a natural temporal signal that may contain temporal trends, seasonality, and noise. The recently developed BFAST (Breaks For Additive Season and Trend) method has been shown successful in detecting breaks from times series of vegetation indices regionally and globally. The principle behind the BFAST algorithm is that it decomposes a time series signal in trend, seasonal and residual components and allows for the detection of abnormal behaviors, called breaks or anomalies. BFAST has been used for moderate spatial resolution data that deliver dense and regular time series, but the method would gain from being applied to higher resolution data as land monitoring requirements tend to move toward higher levels of details. Although BFAST is directly applicable on high resolution data, ongoing investigations using Landsat time series show that fitting a seasonal trend model on these data can be sub-optimal, depending on the temporal density of observations and on the quality of pre-processing applied. As a consequence, while both Landsat and MODIS types of data can be used separately in the break detection process, both have limitations for forest monitoring applications. For medium resolution data like MODIS, the pixel size tends to be larger than the focal scale of most deforestation processes, while the revisit frequency of Landsat data does not allow a good modeling of vegetation seasonal behaviors. Although, the potential of a combined sensor approach has been suggested in earlier studies in order to deal with such space-time trade-offs, this has not been adopted for the BFAST approach yet. The present study proposes, based on the existing BFAST method, to combine Landsat and MODIS for change detection. Such method benefits from the advantages of both sensors, namely the high spatial resolution of Landsat and the high temporal resolution of MODIS. Given that vegetation trends and phenology are mostly regional parameters, the seasonal-trend model required by the BFAST algorithm for defining a stable history period can be fitted using MODIS data. Later, the actual break detection is done at 30m resolution by comparing more recent observations derived from Landsat to the seasonal trend model derived at MODIS scale. The above described method is applied and tested over a tropical dry forest area located in lowland Bolivia. This proof of concept also illustrates a potential of synergies between the upcoming sentinel 2 and 3 sensors for land cover change detection applications.