Automatic Recognition Of Crater-Like Structures In Terrestrial Images J.Earl 1, A.Chicarro 2, C.Koeberl 3, P.G.Marchetti 4, M.Milnes 1 22 nd March 2005 1 LogicaCMG UK (Space & Defence), 2 ESA/ESTEC Scientific Programme, 3 Dept. Geological Sciences, University of Vienna, 4 ESA/ESRIN Earth Observation Programme 1
Overview Our project. Why study Terrestrial Impact Craters? What do Terrestrial Impact Craters look like? Automatic recognition of Impact Craters. Results of the first phase of our study. The BP Crater. Conclusion. Aerial Photograph of Barringer (Meteor) Crater, Arizona. Aerial Photograph of Brent Crater, Canada. 2
Why Study Impact Craters? As a scientific exercise. Distribution of craters on planetary surfaces is correlated to the geological age of the exposed surface area. Link between impact events and mass extinctions. But the distribution of known craters on Earth is significantly different to the distribution on other planetary bodies Mola image covering the surface of Mars between 24 and 40 degrees South, 160 and 173 degrees West. 3
Impact Craters On Earth From A. Chicarro, A. Abels et al. (2003), ERS Synthetic Aperture Radar Imaging of Impact Craters (69 pp), SP-1275, ESA Publications Division, ESTEC, Noordwijk. 4
Why the difference? Earth is more geologically active. Remnants of impact events are eroded, or covered in sediment. Areas where we might expect to find impact structures aren t full of impact crater researchers Which is where an automatic search of remote sensing data comes in. Landsat ETM composite (742) of the Zhamanshin crater in Kazakhstan. 5
Why Study Impact Craters? As a Data Mining exercise. The initial detection of impact crater candidates using remote sensing imagery is a labour intensive process. We gather Terabytes of data each year from multiple sensors. Expert human analysis of all these data is simply not possible. Although confirmation of impact probably isn t possible solely through the use of image mining techniques, we can narrow the search. Multi-Spectral (eg. Landsat) Radar (eg. Envisat ASAR) DEM (eg. SRTM) Detected Crater Candidates 6
What do Impact Craters look like? Depends on age, target material, composition and size of impacting body Simple craters have diameter less than 2-4 km. Complex craters have diameter greater than 2-4 km. Key features: Presence of fractured / shockmetamorphosed bedrock. Generally circular. Simplified cross-sections of typical simple and complex impact crater structures. 7
What do Impact Craters look like? Only the simplest, freshest craters match this morphological template exactly. Brent is a 3.8 km diameter simple crater formed 450 Ma ago. It is vegetated, and filled in by ~250m sedimentary rock. There s little or no crater rim to speak of. 8
What do Impact Craters look like? Oasis is a ~14 km diameter complex structure in Libya. Although less than 150 Ma old, the crater rim has been heavily eroded. The central uplift (~ 4km in diameter) has collapsed. 9
Automatic Recognition. General Approach: Model impact craters as circles. Detect circles on imagery. Use results from multiple images to narrow down number of candidates. Apply expert rules based on expected morphology, land cover etc. Use of Voting rather than Pattern Matching techniques. PCA techniques have been used for matching volcanoes on Venus (Burl et al, 1994) but these don t exhibit the same degree of variability as terrestrial impact craters. 10
Circular Feature Detection The classic way to detect circles in images, particularly impact craters, is to use the Circular Hough Transform. This relies on an edge detection step taking place prior to performing the transform. The resulting parameter space is then classified with a clustering algorithm to find the circles. Various edge detection filters applied to an ASAR WS image of the Aorounga Crater, Chad. Dependent upon quality of edge detection, and choice of parameters. 11
Circular Feature Detection Radial Consistency Algorithm. Models circles as having localised rotational symmetry: the profile through the centre of the feature, taken at various angles, has a degree of consistency that isn t present in a non-circular feature. Results in a 2D parameter space that can be filtered to find the circles. Not reliant on an edge detection step we can match similar spectral characteristics as well as edges. Cross-sections through an SRTM DEM of the Barringer Crater. We see how the crosssection through the centre is more consistent. 12
Circular Feature Detection Example chain of processing leading to detection of the Aorounga crater central uplift. Original image is a Landsat ETM 6a dataset. Creation of parameter space. Filtering of parameter space to reduce contributions from noise / sharpen peaks. Detection of likely circle. 13
Fusion Parameter spaces from multiple input images can be fused removing peaks due to one particular sensing technique. Fusion can take place at various stages (eg. during preprocessing, after creation of parameter space, after crater detection). Landsat Parameter Space SRTM Parameter Space Resulting crater detection 14
Removing False Positives There are many non-impact structures that also exhibit circularity on some images. These can be rejected using their characteristics such as: Non-crater like morphology: No rim. Dimensions. Flooding (to detect lakes). Geology (to detect volcanic structures). We re now looking for other sources to help remove such positives eg. Proba. Detected false positive rejected on morphological grounds Landsat ETM 321, ASAR WS, SRTM. 15
Gross Brukkaros Tall structure of volcanic origin. Can be removed from results using morphology ratio of internal / external height. 16
Our Prototype Based on IDL / ENVI 4.1. Runs in two configurations standalone or batch processing for larger data sets. itool component of the ICDY Prototype Example batch processing chain. 17
Results The following section contains a number of results from the first stage of the ICDY study. 18
Barringer Barringer was the first terrestrial crater identified as being of impact origin. The debris field highlighted in the central image below is still leading to competing theories about how the structure formed. 19
Tswaing Tswaing, in South Africa, is a similar size to Barringer but far more heavily eroded. The rim is vegetated and the crater contains a small salt lake. Photograph of Tswaing from the rim of the crater. 130 x 180 km scene showing Tswaing (top left) and two unfiltered false candidates. 20
Brent We see here how applying additional morphological checks removes some false positives. The 3D projection shows how we pick up the inside of the crater rim though no rim is visible at the SRTM s 90m resolution. 21
Mistastin Mistastin, in Canada, is a much larger complex crater with a diameter of ~ 16 km. The scene shows how we re still able to identify these cases. 22
Other interesting cases Thermal IR channel images are particularly interesting in desert images. These can show changes in rock types. Gosses Bluff (Australia), Haughton (Northern Canada), Aorounga (Chad) imaged using Landsat ETM 6a 23
The BP crater, Request for Proba Data The BP crater in Libya is probably the smallest terrestrial complex impact structure known. Recent ground studies carried out by Professor Koeberl suggest that has a diameter of about 2 km as opposed to the 2.8 km suggested by some remote sensing images. Landsat ETM Band 8 of the BP crater. 24
The BP Crater (cont.) The study is requesting a PROBA image of BP because: It will provide a test case for the use of super-spectral data at an appropriate resolution for crater detection. We already have a range of other remote sensing data to compare it with. BP is small enough to fit in a single PROBA scene. Professor Koeberl has conducted recent systematic insitu studies of the BP crater. Resampled ERS image of BP crater. 25
Conclusions The study has demonstrated that the use of multiple sources of data can significantly improve the accuracy of Earth impact crater detection. The process by which a number of potential candidates is found using a low threshold in a probability space, refined by the application of encoded expert knowledge seems to be a useful compromise when dealing with objects, like impact craters, which exhibit a large degree of variability. The current phase of the study is focussing upon the use of additional data sources (e.g. PROBA, Mars Express) and the processing of a larger area of North Western Africa using Landsat / SRTM / ASAR. 26