6.3 Further research
6.3.2 Detection of rainfall-triggered landslides
surfaces (Collins and Jibson, 2015; Ferrario, 2019; Keefer, 1984; Yamagishi and Yamazaki, 2018). In Nepal, ground failures were primarily a mixture of slides and falls (Collins and Jibson, 2015). In theory, any landslide of sufficient area is expected to disrupt the Earth’s surface, altering its scattering properties and resulting in a temporary coherence loss. However, since I have not attempted to test different types of landslides separately, and my analysis includes very few or no flow-, topple- or lateral-spread-type landslides, it is not certain that all types would have the same signal in coherence-based landslide methods. For methods of landslide detection based on SAR amplitude (e.g. Konishi and Suga, 2018), I would expect that landslides which strongly alter the orientation of a pixel to the SAR sensor by the removal or addition of material would have a stronger signal. Landslides that either alter the material detected by the SAR sensor, for example when vegetation is removed and bare rock is exposed, should also show up strongly in amplitude images, while, for example, a lateral-spread-type landslide may not result in such a strong signal.
The effect that landslide type has on different SAR-based detection is worth in- vestigating for two reasons. First, if a SAR-based detection method is not capable of detecting all landslide types, this must be taken into account when interpreting any output, regardless of whether it is used for research or for emergency response. Second, if different landslide types have different signals in SAR data, it may be possible to draw information on landslide type from SAR, providing a more com- plete understanding of the effect that an earthquake has had on the landscape. In particular, it may become possible to derive information on landslide type us- ing data from increasing numbers of high resolution SAR satellites, such as the recently launched commercial ICEYE SAR constellation, which has an azimuthal resolution of up to 25 cm (ICEYE, 2020). Therefore, while it was beyond the scope if this thesis, I believe it would be beneficial to investigate the SAR coherence and amplitude signals of different types of landslides.
6.3.2 Detection of rainfall-triggered landslides
My focus in this thesis has been on earthquake-triggered landslides. A large pro- portion of SAR-based landslide detection methods, particularly those using SAR coherence have also used earthquakes as case studies (Aimaiti et al., 2019; Ge et al., 2019; Jung and Yun, 2019; Konishi and Suga, 2019; Yun et al., 2015). Those ap- plied to rainfall-triggered landslides often focus on individual large landslides (e.g. Czuchlewski et al., 2003; Mondini et al., 2019) as opposed to large-scale detection of landslides across a landscape, or use polarimetric SAR, which is not acquired after most events (e.g. Czuchlewski et al., 2003; Masato et al., 2020). However,
6.3.2. Detection of rainfall-triggered landslides
the application of large-scale, SAR-based landslide detection methods to rainfall triggered landslides would be advantageous for several reasons. There is an obvi- ous correlation between rainfall and cloud cover, so that optical satellite imagery of rainfall-triggered landslides is likely to be obscured by cloud. Also, although earthquakes can trigger thousands of landslides, they occur relatively rarely, while some areas of the world experience rainfall-triggered landsliding every year. The preference for testing on earthquake-triggered landslides is unsurprising. First, since it is common practice to use InSAR to map ground deformation after the earthquake, SAR data are commonly acquired and processed after an earthquake (NASA, 2018). In some cases, areas of landslides corresponding to areas of low coherence in an interferogram are obvious (e.g. Fujiwara et al., 2019; Vajedian et al., 2018), which would naturally encourage their use as landslide detection case studies. Satellites that are intended for use in emergency response, such as ALOS-2, are likely to acquire a SAR image following an earthquake for use in ground deformation measurements, but historically there has been no need to acquire SAR data after a rainfall event. This has only recently begun to change with the launch of Sentinel-1a in 2014. Currently, the Sentinel-1 satellites acquire either ascending or descending track SAR imagery every 12 days over all land masses globally, and both tracks over tectonic belts (ESA, 2019). This makes it far more likely that SAR data will be available for landslide mapping after a rainfall event. Second, ground shaking due to an earthquake generally lasts for less than a minute, and it can be assumed that landslides occur during, or shortly after this interval, which means that the timing of the landslides is already known, and in a time series of interferograms, we know which one to consider as ’co-event’. The majority of landslide-detection methods, including all of those I have presented here, require this information. They also require all landslides to occur in the same interferogram, which may not be the case, particularly for long rainfall events such as the Nepal monsoon. One solution to this could be to increase the time window of the interferogram, however this has an adverse effect on background coherence, particularly in vegetated regions. I gave an example of this in Chapter 4 when I compared the performance of the same methods on landslides triggered by a single earthquake in the 2018 Lombok earthquake sequence and on landslides triggered by the whole sequence.
A further complication of applying SAR coherence methods to rainfall-triggered landslides is that the rainfall itself can also affect the coherence. It has been demonstrated that rainfall can cause significant decorrelation of an interferogram as wet and dry soil have very different dielectric properties (Nolan and Fatland, 2003; Scott et al., 2017). Theoretically, this could be a significant problem when applying SAR coherence methods to rainfall-triggered landslides. However, the success in