Datasets and pre-processing
Pre- and post-event SPOT-5 images with three spectral bands (green, red, near infrared) and a spatial resolution of 2.5 m (panchromatic) and 10 m (multispectral) were available for the mapping of landslides and landslide changes. The SPOT-5 images illustrate the status quo before and after the Typhoons Aere and Matsa, respectively. After pan-sharpening in ERDAS IMAGINE® (Hexagon Geospatial) software with the Modified IHS Resolution Merge method (Siddiqui 2003) the images were co-registered to each other. Visual inspection after co-registration revealed that in a few areas shifts of about one pixel to two pixels occurred, which is due to the considerable differences in relief and the difficulty to find adequate control points in this natural landscape. Additionally, a DEM with 5 m spatial resolution and its derived products, i.e., slope and curvature, were integrated in the analysis to support the semi-automated classification of landslides and debris flows/sediment transport areas (see Table 1).
Object-based classification based on post-event satellite imagery
The proposed semi-automated object-based approach for the classification of landslides and debris flows/sediment transport areas was developed based on the post-event SPOT-5 image from 2005 and the 5 m DEM including its derivatives. Image analysis was conducted in eCognition® (Trimble) software, defining a set of knowledge-based classification rules.
First, additional layers (Normalized Difference Vegetation Index - NDVI, Green Normalized Difference Vegetation Index - GNDVI, brightness) were calculated. The selection of an appropriate scale parameter for multiresolution segmentation was supported by the Estimation of Scale Parameter 2 (ESP 2) tool, which was implemented as customized algorithm in eCognition software. ESP 2 identifies statistically relevant image object levels for a set of input layers by evaluating the relative changes in local variance for user-defined scale ranges (Drăguţ et al. 2014). A scale parameter of 32 was considered to be optimal for the combined segmentation of the three SPOT-5 bands, as well as the NDVI and brightness layer.
Extraction of landslide candidates
Landslide failures cause significant vegetation loss resulting in a distinct spectral contrast between landslides and their surroundings, supporting the detection of landslides in optical imagery (Behling et al. 2014). Such contrasts can especially be observed for event-based landslides occurring in densely vegetated regions such as Taiwan. The absence of vegetation and the presence of bare ground were assumed to be an indication of potential landslide areas. Therefore, low NDVI values of image objects, relative to the information contained in the image, were used to identify landslide-affected areas. The initial NDVI-based extraction of landslide candidates was slightly refined by using brightness in combination with context information. For instance, small shadow areas adjacent to landslide candidates were included in the classification. The extracted candidate areas were further divided into two classes: shallow landslides and debris flows/sediment transport areas.
Landslide classification and class refinement
Due to spectral similarity debris flows/sediment transport areas were mainly differentiated from shallow landslides by using morphological characteristics of image objects. An independent segmentation of plan curvature and slope with a scale parameter of 15 was performed within a sub-project in eCognition to facilitate the delineation of channel-like features where debris flows/sediment transport areas likely occur. Working with such sub-projects allows independent processing (e.g., segmentation, classification) for the same area. In the end, results can be merged while preserving the initial image object boundaries. First, objects with curvature values < 0 were classified as concave terrain where sediment is most likely accumulated and transported downstream. Next, texture and context information of image objects were used to enhance the extraction of stream channels, particularly to remove several false positive objects fulfilling the curvature threshold. GLCM (Gray Level Co-occurrence Matrix) contrast of the slope layer, which measures the local variations in the GLCM, was applied to remove already classified objects with a GLCM contrast (all dir.) value < 0.15. Directional GLCMs, especially the GLCM contrast between landslide areas and surrounding areas, appear to be very efficient in landslide detection (Blaschke et al. 2014b). The thresholds were selected based on visual assessment. Finally, small objects which were isolated, and thus, unconnected to channels were removed from the classification. The classified debris flows/sediment transport areas were then merged with the previously classified landslide candidates. Areas where both classes overlapped were further treated as debris flows/sediment transport areas. This procedure allowed classifying debris flows/sediment transport areas by considering spectral characteristics, as well as morphological characteristics derived from a spectrally independent segmentation based on DEM derivatives. It is noteworthy that the class debris flows/sediment transport areas also includes the river beds where the downstream transportation of debris and sediments occurs. The remaining landslide candidates were classified as shallow landslides. Subsequently, the two classes were iteratively refined by using spatial (e.g., area, shape) and contextual parameters (e.g., relative border to neighbouring objects) of image objects to eliminate false positives with spectral properties similar to landslide areas (e.g., cleared bamboo forests, paths, harvested agricultural fields). The boundaries of the classified image objects were finally smoothed with growing and shrinking operations.
Testing the transferability of the classification approach
The described approach (cf. Fig. 3) was transferred to the post-event SPOT-5 image from 2004, which showed radiometric differences compared to the image from 2005 (see Table 2). More shadows were present, and the vegetation appeared to be darker. Nevertheless, since the usage of absolute spectral thresholds was minimised, the developed OBIA classification approach could be transferred with only minor adaptations to the 2004 post-event image.
The two landslide classifications based on the 2004 and 2005 post-event images served as input for the class-specific landslide change detection as described in the following section.
Object-based change detection
The class-specific object-based change detection was initially accomplished for the Typhoon Matsa event (2005) and then transferred to the images available for the Typhoon Aere (2004). Fig. 3 gives an overview about the developed workflow for landslide classification and landslide change mapping.
Multi-temporal image segmentation
First, a joint multiresolution segmentation on the 2005 pre- and post-event images, denoted as “image pair”, was performed using the three bands of the SPOT-5 images and the NDVI layers, in total eight bands. Such a multi-temporal object change detection approach produces spatially corresponding change objects, since temporally consecutive images are combined and segmented together (Chen et al. 2012). Again, the selection of an appropriate scale parameter, in this case 31, was supported by the statistical evaluation with the ESP 2 tool. Each object of the segmentation result was then analysed with respect to its transformation between the pre- and post-event image.
NDVI and GNDVI normalisation
New landslides and debris flows/sediment transport areas were recognised by comparing NDVI and GNDVI values of segmentation-derived image objects between the corresponding pre- and post-event images. Vegetation loss was identified by a negative change of the NDVI and the GNDVI. As the radiometric characteristics between the image pairs slightly differed (different appearance of vegetation and shadows mainly because of different acquisition dates during the year; see Table 1), the comparison of the absolute NDVI and GNDVI values would have been biased. Therefore, the vegetation indices across the respective pre- and post-event images were normalised by using normalisation factors. The normalisation factors were calculated by dividing the mean value of the pixels of the post-event NDVI and GNDVI, respectively, by the value of the corresponding pre-event index. By multiplying the pre-event NDVI and GNDVI with the corresponding normalisation factor, the indices were calibrated to the radiometric characteristics of the post-event image. Consequently, the comparability of the pre- and post-event indices was increased.
For each image object, the difference in the NDVI and GNDVI was calculated by subtracting the post-event value from the normalised pre-event value. For the NDVI, if the change was higher than a defined threshold of 0.5, the objects were classified as affected by change. For the GNDVI, the change threshold was set to 0.4. These thresholds were manually determined after visual assessment and used to prevent the classification of objects, which showed only minor value changes. In each case, a second condition was applied to detect only those objects with a relatively low NDVI or GNDVI value in terms of the relative information contained in the post-event image, i.e., objects with very little or no vegetation. Hence, areas which had been affected by a change in vegetation, but which were still vegetated, could be excluded. Most of the changes were recognized by applying the NDVI; the GNDVI allowed for the detection of several small-scale changes, which were not captured by the NDVI. Finally, the objects affected by change were synchronised, i.e., merged, with the prior established classification of shallow landslides and debris flow/sediment transport areas, and thus, changes were attributed to landslide classes. Small sliver polygons resulting from the synchronisation were removed by applying object growing and shrinking algorithms. The same procedure was applied for detecting the changes caused by Typhoon Aere, whereby no adaptations in the ruleset were necessary.