ALADIM-HR: Automatic LAndslide Detection and Inventory Mapping from multispectral HR (S2 or L8) data

../_images/tuto_aladim_icon.png

ALADIM

This service is developped by CNRS-EOST (Strasbourg, France). It allows to detect and map new landslides triggered by large forcing events (earthquake, heavy rains) from the analysis of pre- and post-event imagery, and is based on change detection methods. It allows the processing of High Resolution multispectral data (ALADIM-HR; Sentinel-2 or Landsat-8) and Very-High Resolution multispectral data (ALADIM-VHR; typically Pléiades and Spot 6/7). The set of pre- and post-image should be accurately co-registered in order to use the service. A training dataset of manually mapped landslides (by digitalization), the extent of the training areas, and the extent of the region of interest (ROI) should be provided as inputs (shape file-format) by the user. The outputs consist in a database of landslide polygons than can be assimilated to an Earth-Observation derived landslide inventory. ALADIM builds on the change detection methodology partially described in [1] and [2] .

EO sources supported:

  • Sentinel-2 MSI LIC (SAFE file format),
  • Landsat-8

Input specifications

Beside the service parameters an archive folder containing the training set, the training areas (and aoi) in shapefile format is needed. See the tutorial (tutorial) to create these inputs.

Output specifications

  • A shapefile (*.shp files) containing the landslides detected at an F2 optimal threshold.
  • An image (geotiff file format) containing all landslides detected at an F2 optimal threshold.
  • Two documents (*.pdf files) presenting the cross-validation quality control (precision-recall curves and acurracies of the parameters).

This tutorial introduces to the use of the service for the detection and the mapping of landslides from HR multispectral images. To this end, we will process a couple of Sentinel-2 images acquired before and after the `Idai Cyclone`_ which hit Mozambique on 14 March 2019.

Select the processing service

  • Login to the platform (see user section)
  • Go to the Geobrowser, expand the panel “Processing services” on the right hand side and select the processing service “ALADIM-HR”:
../_images/aladim_tuto_img1.png

This will display the service panel including several pre-defined parameters which can be adapted.

../_images/aladim_tuto_img2.png

Use case: Landslide detection and mapping from HR multispectral data

Select input data

The Geobrowser offers multiple ways to search Sentinel 2 dataset with spatial and temporal filters. The interested reader should refer to the Geobrowser section for a general introduction. For this tutorial we will show the example of a research of a pair of Sentinel 2 images which encompass the area of interest around Chimanimani (Mozambique). The first image was search before the Cyclone and the second after the event.

Select Sentinel-2 from the EO Data pulldown menu:

../_images/aladim_tuto_img3.png

Draw a polygon on the map around your area of interest and reduce the time extend thanks to the timeline at the bottom of the map:

../_images/aladim_tuto_img4.png

Drag and drop the images of your choice in the fields of the service panel associated with the pre-event and the post-event Sentinel-2 images:

../_images/aladim_tuto_img5.png
../_images/aladim_tuto_img6.png

Create an archive for the ensemble of your input shapefiles (training_areas.shp, training_samples.shp and aoi.shp). The framework requires a flat .tar.gz format (i.e. the contents of the archive file must not include folders). A tutorial about the input dataset creation can be found here tutorial

Upload the archive:

../_images/aladim_tuto_img7.png
../_images/aladim_tuto_img8.png
../_images/aladim_tuto_img9.png

Drop the archive in the field of the service panel named “shapes files uri”:

../_images/aladim_tuto_img10.png

Set the processing parameters

There is a total of 5 processing parameters that can be adjusted. When hovering over the parameter fields, you will see a short explanation for each of the parameters.

  • ALADIM_SEG_SCALE: The segmentation scale factor (See [3] for details about segmentation). Larger values will result in fewer larger segments and faster processing. Smaller values will result in more more small segments which will increase the processing time but also typically the accuracy of the classification. The default value is 70 but the value depends a lot on the value range of the input imagery and the landscape characteristics.
  • ALADIM_SEG_COLOR_WEIGHT: A value between 0 and 1 to define the weight of color during the segmentation. The default value is 0.9.
  • ALADIM_SEG_MIN_SIZE: Minimum allowed segment size. Segments smaller that this value (in pixels) will be merged to their most similar neighbor after the segmentation or deleted if isolated.
  • ALADIM_POSITIVE_FRACTION: A value between 0 and 1. If the fraction of positive area (i.e. landslide as mapped in the training samples) within a segment exceeds this value it is considered as a positive example. Vice versa it will be considered as a negative example. The default value is 0.5.
  • ALADIM_USE_CLOUD_MASK: If set to True the FMASK algorithm [5] will be used to detect clouds, snow, and water and mask them from the segmentation.

The figure below summarizes the parameter settings for this test.

../_images/aladim_tuto_img11.png

Run the job

  • You are good to go. Click on the button Run Job at the bottom of the right panel. Depending on the allocated resources the execution will require a few hours to terminate.
../_images/aladim_tuto_img12.png
  • Once the job has finished, click on the Show results button to get a list and a pre-visualization of the results.

Note

The pre-visualization in the Geobrowser is just a preview and the user is encouraged to download the results for further analysis and post-processing.

../_images/aladim_tuto_img13.png

References

[1]Stumppf, A., Kerle, N. 20110. Object-oriented mapping of landslides using Random Forests. Remote Sensing of Environment, 115(10): 2564-2577.
[2]Stumpf, A., Lachiche, N., Malet, J.-P., Puissant, A., Kerle, N. 2014. Active learning in the spatial domain for remote sensing image classification. IEEE Transactions on Geoscience and Remote Sensing, 52(5): 2492-2507.
[3]Lassalle, P., Inglada, J. Michel, J., Grizonnet, M., Malik, P. 2015. A scalable tile-based framework for region-merging segmentation. IEEE Transactions on Geoscience and Remote Sensing, 53(10): 5473-5485.
[4]Walvoort, D.J.J., Brus, D.J., De Gruijter, J.J. 2010. A R package for spatial coverage sampling and random sampling from compact geographical strata by k-means. Computers & Geosciences, 36(10): 1261-1267.
[5]Zhu, Z., Wang, S., Woodcock, C.E. 2015. Improvement and expansion of the Fmask algorithm: cloud, cloud shadow, and snow detection for Landsats 4-7, 8, and Sentinel 2 images. Remote Sensing of Environment, 159: 269-277.