Kyoto TerraSar X Work

Extracted procedural details from "Automatic Detection of Landslides from SAR Images: Application to the 2011 Kii Landslides", Masumi Yamada, Manabu Hashimoto, Yo Fukushima, Yuki Matsushi, Masahiro Chigira (APSAR2013 Young Scientist Award winning paper)

Data: "TerraSAR-X [...] an active phased array X-band SAR antenna (wavelength 31 mm). [....] descending tracks with the stripmap mode with HH polarization. [...] about 0.9 ~ 1.4 m and 1.8 ~ 2.0 m resolutions in range and azimuth directions [...] only coregistered and geocoded intensity images. ASTER Global Digital Elevation Model (GDEM) ver.2 were used for geocoding. [...] converted intensity images into Sunraster files."

Criteria: "landslides with area of 1 ha or larger were extracted from aerial photos and the results of field surveys."

Method: "Due to the steep topography and forest vegetation [..], the image pairs did not have enough coherence for interferometry. [...]. We composed a single image by assigning intensities of the pre-event image to a red channel and those of the post-event image to green and blue (cyan) [...and were] able to identify more than 20 large landslides [...] visually. For the automatic detection [...] suppress noise [...], identify signals, and classify landslides from other surface changes."

A. Suppress noise, mask foreshortening zones

1. Noise: "bilateral filter [8]"
2. Foreshortening: "Pixels with intensity of outside of the measurement range ... were masked"

B. Identify signals

"The roughness change ... from forest to rock surface makes the backscattering generally larger. The change in the angle between the incidence direction of the wave and the ground surface also changes the intensity of the back-scattered wave."

C. Distinguish landslides from other events

"... deposits of debris flow, water level change of a river, and cutting of forest trees [etc.] change roughness of ground surface ..."

1. Histogrammed color distribution
2. "... removed the objects on which one color occupied a certain percent of pixels."

D. Optimizing the thresholds

Parameters

... for filter control:

• "standard deviation of the bilateral filter”
• "threshold of intensity for masking
• "window size of the median filter."

... for signal identification:

• "threshold for binarization"
• "size of the morphological closing and opening"
• "threshold for the area"
• "threshold of percentage of pixels with red or cyan"

"... 8 parameters ..." (including "certain percent" in C, above?)

"... define the object as a landslide if the difference of the location is less than 0.7 km.

"320 trials" to identify optimal parameters.