Comparing object-based landslide detection methods based on polarimetric SAR and optical satellite imagery – a case study in Taiwan

Abstract

Applied to a test site located in southern Taiwan, this study compares two object-based image analysis (OBIA) methods for post-failure landslide detection based on (I) polarimetric synthetic aperture radar (PolSAR) and (II) optical satellite imagery. With its day-and-night availability and almost complete weather independency, SAR has several advantages compared to optical imagery. Consequently, in most cases, SAR imagery for a dedicated area of interest is earlier available than the first cloud-free optical data. However, the high spatial and spectral resolution of multispectral optical Earth observation data may enable a more detailed and accurate landslide detection. A comparison of both methods is feasible, as the polarimetric SAR image (dual-polarimetric (HH/HV) TerraSAR-X, StripMap) and the very high spatial resolution optical imagery (QuickBird) were acquired at a temporal baseline of only 20 days. Thus, SAR and optical data show the same state of the environment. After speckle filtering of the PolSAR data using the refined Lee filter and radiometric calibration, the intensity information of both polarization channels (HH and HV) was geocoded. Next, the features of interest (i.e. landslides, debris flows and riverbed) are derived from the PolSAR imagery using a newly developed OBIA procedure, which makes use of the different backscattering behavior of forest and other vegetated areas as compared to bare soil, which was assumed to be an indication for the occurrence of mass-movements and debris/sediment transport and deposition areas. Using the normalized difference standard deviation of the calibrated intensities of both polarimetric channels, HH and HV, the OBIA procedure considers (I) the higher variation of the backscattering intensities in forest areas and (II) the relatively higher backscattering of vegetated areas in the cross-polarized channel compared to bare soil areas. The latter is characterized by a more dominant backscattering in the co-polarized channel. For the object-based detection based on the QuickBird image the Normalized Difference Vegetation Index (NDVI) was applied to detect the unvegetated areas. Based on the NDVI layer an automated threshold was computed, which divides the image into two subsets, i.e. vegetated and non-vegetated areas. To produce suitable image objects for the classification, the areas potentially affected by mass-movements were re-segmented using the multiresolution segmentation algorithm implemented in the eCognition (Trimble) software. Additionally to the QuickBird image a digital elevation model (DEM) with 5 m spatial resolution was used to support the differentiation of classes. As the spectral information alone was not sufficient for class separation, the distinction into landslides, debris flows and riverbed was mainly based on slope values. Finally, a few rules considering spatial and contextual properties were introduced to refine the classification and to remove false positives (e.g. built-up areas and fields). The reference dataset for validation includes vector data of landslides, debris flows and the river bed and was produced through manual digitization, performed by a local expert. Both results are compared to the reference data set and the pros and cons of the imagery utilized for landslide detection are evaluated. The result of the novel object-based method based on PolSAR data reveals a certain potential for landslide detection, especially for rapid assessment of affected areas after landslide triggering events.