Spatiotemporal ICA/PCA decomposition of optical displacement field stacks: perspective for landslide time series inversion

Monitoring ground surface motion is a key information to locate active landslides and possibly detect failure onsets but also to better understand their mechanical behavior in relation with environmental forcing. In-situ and remote technologies are available to provide measures of the ground displacement with different advantages and limitations (in terms of spatial coverage, sampling frequency, etc.). Image matching techniques have been commonly used to detect and measure landslide acceleration but this is often limited to a small amount of images. In the recent years, the number of optical satellite constellations have significantly increased providing global coverage with a frequent revisit time at medium to high spatial resolution and an open access policy (e.g. Sentinel 2, Landsat 7/8). These datasets present new perspectives for the monitoring of slow (cm/day) to moderate (m/month) landslide motion and poses challenges to discriminate between the different spatio-temporal sources (e.g. rainfall correlated signal, noise, seasonal signal, etc.) present in the time -series. 

We investigate the use of spatiotemporal ICA/PCA decomposition on optical displacement stacks of landslide areas. The main goal aims at testing 1) the capability of ICA/PCA analysis to detect relevant deformation deformation sources in the case of landslide monitoring and 2) the possibility to improve the time-series inversion of landslide motion by removing spatiotemporal sources that can result from seasonal sun exposition or geometric inaccuracies. We use the MPIC-OPT-Slide service of the GeoHazards Exploitation Platform (GEP) to compute several correlograms and displacement fields (>500 per site) from Sentinel-2 acquisitions on the slow-moving La Valette landslide (Alpes-de-Haute-Provence, France) and the moderately-moving Aiguilles-Pas de l’Ours landslide (Hautes-Alpes, France). We show that in case of steady-state deformation, the noise can be significantly removed around the active parts of the slope. In the case of more complex deformation evolution, pertinent sources can be manually isolated but the choice of the number of sources and their automatic selection remain challenging.