The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Publications Copernicus
Articles | Volume XLI-B7
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLI-B7, 987–989, 2016
Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci., XLI-B7, 987–989, 2016

  14 Oct 2016

14 Oct 2016


Meng Lu1 and Eliakim Hamunyela2 Meng Lu and Eliakim Hamunyela
  • 1Institute for Geoinformatics, Westfälische Wilhelms-Universität Münster (WWU), Heisenbergstraße 2, 48149 Münster, Germany
  • 2Laboratory of Geo-Information Science and Remote Sensing, Wageningen University, Droevendaalsesteeg 3, Wageningen 6708 PB, the Netherlands

Keywords: Multi-Spectral, BFAST, Dimension Reduction, Deforestation Monitor

Abstract. In recent years, the methods for detecting structural changes in time series have been adapted for forest disturbance monitoring using satellite data. The BFAST (Breaks For Additive Season and Trend) Monitor framework, which detects forest cover disturbances from satellite image time series based on empirical fluctuation tests, is particularly used for near real-time deforestation monitoring, and it has been shown to be robust in detecting forest disturbances. Typically, a vegetation index that is transformed from spectral bands into feature space (e.g. normalised difference vegetation index (NDVI)) is used as input for BFAST Monitor. However, using a vegetation index for deforestation monitoring is a major limitation because it is difficult to separate deforestation from multiple seasonality effects, noise, and other forest disturbance. In this study, we address such limitation by exploiting the multi-spectral band of satellite data. To demonstrate our approach, we carried out a case study in a deciduous tropical forest in Bolivia, South America. We reduce the dimensionality from spectral bands, space and time with projective methods particularly the Principal Component Analysis (PCA), resulting in a new index that is more suitable for change monitoring. Our results show significantly improved temporal delay in deforestation detection. With our approach, we achieved a median temporal lag of 6 observations, which was significantly shorter than the temporal lags from conventional approaches (14 to 21 observations).