PlanetScope time series for the operational monitoring of large landslide terrain motion.

PlanetScope imagery, with its high spatial resolution (3 m) and high revisit time (possibly 1 day, according to cloud cover) is a game changer for operational landslide monitoring especially for monitoring surface deformation using Optical Image Correlation (OIC) approaches. The high spatial resolution should allow to enhance both the sensitivity and the accuracy of the measurements with the possibility to obtain a theoretical deformation detection of 0.30 m. The high revisiting time ensures the completion of dense image time series, useful to increase the Signal-to-Noise ratio associated with multiple image pairing and to possibly construct deformation time series on daily temporal scales. These two aspects of PlanetScope imagery fill the gap of current optical constellations that usually offer either lower spatial resolution with regular and rather short revisit time (e.g. Sentinel-2, Landsat-8) or very high spatial resolution with irregular revisit time (e.g. Pléiades, Worldview). However, the specifications of the PlanetScope L3B data products do not meet the expected quality in terms of ortho-rectification and image time series co-registration and a specific workflow needs to be implemented. 

We propose a new workflow for -processing  PlanetScope L3B data products. The developed approach consists firstly in removing  clouds and water, using Fmask algorithm and PlanetScope Unusable Data Mask products delivered with the L3B products.

Secondly, we observe that the misalignment between scenes  can go up to 8 pixels of difference (24 m on ground), varying highly within the images and from one image to another. To correct such errors, a co-registration process in two steps is applied. At first, using the AROSICS library (Scheffler et al., 2017), the misalignment errors at a local scale are computed by image correlation in the frequency domain on overlapping subwindows pinned on a grid covering the whole image. These offsets are used to correct the local scale co-registration errors. After this step, a global shift is still observed between scenes, leading to the second step of co-registration at global scale. The global shift is corresponding to the mean offsets between image tie points, and is corrected by applying these offsets directly on the product. These developments have been integrated within the  GDM-OPT-Slide service and have been tested on two sites: La Valette (South East France) and Aiguilles (South East France) landslides to retrieve the mean velocity and the ground displacement time series for each pixel. We validate the proposed workflow by comparing the results of the processing chain and in-situ dataset (GNSS, Lidar and photogrammetry). We show that the proposed methodology allows envisaging the operational use of Planetscope imagery to document and monitor the displacement of large landslides with velocity larger than 0.3 m/year.