Needles in a Haystack – Trying to Find Landslides in Peru

As part of a larger NSF grant (award #0928526) investigating the M8.0 Pisco Peru Earthquake, I worked at CAST (The Center for Advanced Spatial Technologies) on a side project using remotely sensed imagery donated by the National Geospatial Intelligence Agency to try and locate landslides caused by the earthquake. The goal was to use change detection and automated image analysis in the form of a false-positive filter to find landslides.

We had nearly 900 gigabytes of imagery for the area from a myriad of sensor systems including Quickbird and Worldview very-high-resolution sensors. The first step of the project was simply sorting through the flood of data, for which we used MetaGETA, which I highly recommend.

Image preprocessing included orthorectification using an ASTER 30m x 30m DEM, geometric registration, and radiometric normalization. Change detection was performed in ERDAS Imagine using image differencing with area, major axis, and elongation filtering.

This diagram shows the criteria and thresholds used in the landslide false-positive filter.

This diagram shows the criteria and thresholds used in the landslide false-positive filter.

The blue polygons are landslide candidates, and the red areas are identified by the filter as potential landslide locations. Notice on the combined panel (right) how many blue polygons fall outside the suitable landslide area, indicating a very large number of false-positives that the filter correctly removed.

The blue polygons are landslide candidates, and the red areas are identified by the filter as potential landslide locations. Notice on the combined panel (right) how many blue polygons fall outside the suitable landslide area, indicating a very large number of false-positives that the filter correctly removed.

The false-positive filter was a knowledge-based classification adapted from Martha et al., 2009, implemented in Imagine. The filter resulted in a binary mask that was applied to landslide candidates from the image differencing to filter out false-positives. Our preliminary results showed the filter correctly removed things like roads, buildings, agricultural fields, and other false-positive changes from the landslide candidates.

Ultimately my part of the project was not completed as we never received the data from the field teams to validate our approach beyond a few sample points. Our early results were promising though, and the false-positive filter technique showed a lot of potential. I presented our preliminary results at AAG 2012, as well as a few smaller meetings (see CV for full list).

Advertisements