Dynamic Landmarking

Detecting Transient Surface Features with Dynamic Landmarking


We have developed methods to dynamically and autonomously detect transient surface features, such as dust devil tracks or dark slope streaks on Mars, from images. Most prior work on this subject has relied on manual examination of image pairs. Exciting discoveries of new surface features such as gullies and impact craters have been made, usually serendipitously. How many more such features remain undiscovered in the massive volume of images being collected and returned?

Automated methods can help reduce the manual effort needed to find and catalog new and interesting features. Previous techniques for automated analysis have focused on changes at the pixel level. They require an initial, sometimes slow, full registration between a candidate pair of images. Once the images are registered, subtracting one from the other yields changes. These are usually thresholded or subjected to further analysis to help filter out noise and other uninteresting "changes".

Dynamic Landmarking Approach

In contrast, our approach focuses on the image content, not just the pixels. We first analyze each image to identify visually salient "landmarks", and then compare the detected landmarks between the images to highlight changes. No image registration is required. We proceed as follows:

detection and comparison process
  1. Detect landmarks: Using statistical measures of salience, create a "salience map" that indicates, for each pixel, how salient it is with respect to its local region. Automatically select an appropriate salience threshold, and use it to produce contours around regions of high salience. Each such region becomes a "landmark." Landmarks may include craters, volcanoes, fissures, and so on, as well as the transient features we seek.
  2. Extract features: Compute descriptive attributes for each landmark, such as size (surface area), shape, albedo, homogeneity, etc. Using a trained landmark classifier, assign each discovered landmark as one of a set of known classes of interest (craters, volcanoes, dust devil tracks, gullies, etc.). Mark any unclassified landmarks as new, potentially high-interest regions.
  3. Create a Relative Landmark Graph (RLG): Landmarks (and their features) are the nodes, and edges connect each landmark to its k nearest neighbors.
  4. Match the Graphs and Detect Changes: Compare the landmark sets discovered in two images to quickly detect any changes with high precision. Using the Munkres/Hungarian algorithm, compute a matching between the RLGs from two images. Mark unmatched landmarks as changes.
We have applied this approach to images collected by the Mars Orbiter Camera (MOC), Thermal Emission Imaging System (THEMIS), and other Mars-orbiting spacecraft.


  • Because the landmarks are represented at an abstract level, it is possible to combine observations from different instruments at varying attitudes and under different illumination conditions.
  • Dynamic landmarking can detect transient features without requiring image registration, which represents a large step towards enabling the onboard use of this technology.
  • These landmarks provide a regional characterization of the area covered by the image that can be used to better understand surface processes as well as to recognize when two images overlap.
  • One of the biggest benefits of this effort is the increased productivity it can lend to science investigations, as compared with manual change detection. Our goal is to increase our knowledge about other planetary surfaces, and the dynamic processes present, to support future human and robotic exploration.

Funded by the NASA Applied Information Systems Research Program, 2007-2010.