Loading...
 
ESA > Join & Share > Technology Projects > ICDY Project
Print

ICDY Project

Project Title   Impact Crater Discovery
Project Acronym   ICDY
Contractor(s)   LogicaCMG (UK), Universitat Wien (A)

 

Project Context
          Objectives
             Architecture
          How it Works
Output

 


Context   Top

Impact cratering is increasingly being recognised as one of the driving mechanisms of planetary geological evolution due to the effect of major impacts upon subsequent tectonic events. There are also a number of theories which can be used to estimate the age of regions of Mars by examining the impact crater flux. The enumeration and analysis of such impact craters can thus tell us about the history of both the individual planet and the Solar System.

The process of recognising a new impact crater structure on the Earth falls into two steps:

  Image The detection of a candidate impact site. Many established structures have initially attracted attention because of their appearance as anomalous circular features in topography, or in the bedrock geology.
  Image The verification of the site as a true impact crater by establishing the presence of shock metamorphic effects in the structure's rocks.

The distribution of established impact craters on Earth is therefore at least partially dependent upon the quality of information available to be drawn upon for the detection phase, and upon the accessibility of the candidate sites so that the verification step can be performed. The large archives of remote sensing data (radar, optical, multispectral etc.) can aid the search for impact craters through the application of image processing techniques, but an efficient search requires a degree of automation.

The automation of the analysis of remote sensing data is becoming an issue because of the proliferation of both earth observation and interplanetary multi-sensor missions. The vast quantities of data returned by such missions mean that unassisted analysis is impossible and that any techniques which require a substantial amount of expert human input are both impractical and expensive. This means that automation of this analysis, the classification of remote sensing data, is a desirable goal

Current techniques to deal with the problem of retrieving particular sets of remote sensing data rely on parameters such as sensor, time and location - and not upon the actual content of the data. Future systems should instead allow searches over indices within which details of the data's content are available. The creation of such indices will require the development of a set of techniques and algorithms for the automated recognition of features within the data. Ideally such indices should deal with multiple sources of data, rather than each source individually.

The history of the study of cratering, particularly upon Mars, demonstrates some of the advantages of such a multi-sensor approach. A Viking Orbiter image of the surface of Mars provides a certain amount of information about the topography and distribution of any craters; but combining the results with those found from Mars Global Surveyor laser altimeter map leads to a huge range of additional information becoming available.

The purpose of the Impact Crater Discovery project is to apply a multi-sensor approach to the automatic detection of impact craters on the Earth, and develop a prototype enabling the efficient comparison of alternative algorithms.

 


Objectives   Top

Currently about 160 impact craters have been identified on the Earth - a substantial number of which are not visible on the surface. The craters are distributed worldwide (see Figure 1) but there are significant concentrations of large craters in the Baltic and Canadian shields. Comparisons with other planetary bodies suggest that many more structures are likely to exist in the less well-studied shields of Africa, Asia, South America and Australia, and, of course, in the oceans . On the other hand smaller and more recent craters (less than a few million years old) are best preserved in desert areas of the world.

Distribution

Figure 1. Impact Crater Distribution 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.

Earth impact craters can be classified into two major morphological sets, roughly corresponding to size:

Image Simple impact craters are typically bowl shaped, and have a diameter of less than 4 km. The Barringer crater in Arizona is a particularly clear example of an Earth simple crater. 3D

Figure 2. Landsat 432 image of Barringer Crater, Arizona, USA mapped onto SRTM dataset.

Image Complex impact craters feature a central uplift, as well as an outer rim. Their diameter is usually greater than 2 km. The Aorounga crater in Chad is an example of an Earth complex crater. The uplift / rim structure is clear in radar imagery. ASAR

Figure 3. ASAR WS image of Aorounga Crater, Chad (75 metre/pixel).

Unfortunately, the weathering processes on the Earth mean that few of the remaining known impact structures are as clear as these two; many impact structures have been heavily eroded (e.g. the Gosses Bluff central uplift in Figure 4 has collapsed, whilst the rim has worn away, leading it to look like a simple crater from orbit) and any algorithm developed to automatically detect such features has to take these possibilities into account.

Landsat

Figure 4. Landsat 432 image of Gosses Bluff, Australia

 


Architecture   Top

One of the principal outputs of the ICDY study is the prototype, which is written in IDL and uses the functionality of ENVI. The purpose of the prototype is to allow the rapid prototyping of circular feature detection and crater-classification algorithms. The program is designed so that it can be run in two distinct modes:

Image A fully-featured mode for use with a full ENVI/IDL licence. This version is able to manipulate multiple datasets from a wide range of different data sources and run in a batch processing mode to allow longer, chained, runs of algorithms
 
Image A standalone mode that can be run on any Windows PC without the need for a full ENVI/IDL licence. The processing algorithms in this mode are fixed for a given release when run without the full licence, and the software can only deal with single datasets. However the standalone software provides a convenient environment for testing new algorithm variations quickly on small regions of data, and can be used to review the results from the fully-featured mode.

The use of ENVI as the base for the software means that a wide variety of remote sensing datasets can be swiftly imported into the prototype framework and pre-processed using the included techniques.

Outputs from the prototype include:

ENVI evf files which can superimposed on top of the original imagery, and whose contents can be filtered using simple Boolean logic to determine additional filtering rules that may prove useful.

Envi

Figure 5. Envi prototype analysis.

An ICDY specific .xml file which contains a full list of the properties for a given crater candidate. This .xml file can be loaded directly into the standalone iTool, which allows the user to further filter and manipulate the results.

iTool

Figure 6. iTool prototype analysis.

Additionally reports can be generated as standard webpages, allowing a user to drill through the results over a whole sequence of batch runs down to the original imagery used for a individual crater candidate.


How it Works   Top

Earth impact craters appear in all states of degradation, in all types of rocks, and in all types of regional environments. This means that there is no clear cut answer to the question of what the best channels / wavelengths / techniques are to use when searching for impact structures. All techniques and wavelengths have some merit for the analysis of geological features and hence need to be used in their own way..

The fact that impact craters can appear in any type of rock means that compositional studies at the spatial and spectral resolution afforded by remote sensing do not help search for impact craters by default. Remote compositional investigations are advantageous only in areas that are difficult to investigate in situ, but the knowledge to be gained from such investigations is limited relative to that that can be gained from a proper ground study.

A variety of techniques have previously been used to detect crater like objects on other planetary bodies including both voting (Hough Transform based, Tangent/Curve estimation based on shadows etc.,) and pattern matching (Principal Component analysis based, etc.,) techniques. Pattern matching techniques are not especially suitable for dealing with Earth impact craters because the available training set is small (160 craters) and the variation within that training set is extremely large (in age, size, shape, vegetation cover, erosion level, etc.,).

The variation within the known crater set means that concentrating an algorithm design too heavily upon specific crater morphology can be misleading, since similarly-sized craters on the Earth's surface can exhibit contrasting characteristics (eg. Zhamanshin and Aorounga, or Barringer and Tswaing). For our purposes we begin by characterizing impact craters in a very simple way by using their circular shape as the main feature. Although this is an oversimplification, it provides a first order approximation which allows a given algorithm to operate on data obtained from different sensors under a range of conditions. The algorithm can then be tailored to specific data types (e.g., DEM).

Other algorithms and approaches (e.g. the Circular Hough Transform) were considered and implemented as part of the study.

Circular Feature Detection - The Radial Consistency Algorithm.

The Radial Consistency Algorithm developed as part of the study models impact craters as having localised rotational symmetry (referred to as radial consistency). This means that the profile, taken through the centre of a circular feature at different angles has a degree of consistency (due to the circular symmetry) that is not present in points that are not at the centre of a circular feature. This is easiest to visualise by considering a digital terrain model as demonstrated in Figure 7.

Figure 7 shows this consistency effect using cross-sections through a point in the centre of the Barringer crater, and a point just outside of the crater. The red "star" over-plotted on each image shows the cross-sections sampled. The cross-sections through the centre of the crater are fairly similar, and the cross-sections outside of the crater exhibit a greater degree of variation.

Cross

Figure 7. Cross sections through the Barringer crater.

The Radial Consistency algorithm can be related to the Circular Hough Transform by replacing the test that each pixel (x,y) lies on the circle defined by the triple {a,b,r} with the test that the pixels lies within a region of circular symmetry centred at (a,b). Consequently the peaks in the parameter space {a,b} correspond to the most likely locations of the regions of Radial Consistency in the input image.

The key difference between the Radial Consistency and the Hough Transform algorithms is therefore that the Radial Consistency algorithm is not applied to an edge enhanced image - since this edge enhanced image may well be lacking the very circular patterns the algorithm is attempting to pick up. Further advantages of the Radial Consistency algorithm include the greater ease with the output 2 dimensional parameter set can be visualised (as a 2d image) and combined with additional sources, and the way with which the algorithm scales with increasing radius with order O(R), as opposed to order O(R2) with a standard Circular Hough Transform.

Data Fusion.

One of the key areas of focus for this study was the issue of dealing with multiple datasets, and the use of data fusion techniques to improve the detection process. The reasoning behind this was that images of the Earth have a high complexity - which makes recognition of all but the most obvious craters difficult when using observations made by a single sensor.

In the results presented in this study the fusion was performed in the information space by linearly combining the Radial Consistency parameter spaces before beginning the feature peak detection process. This led to a significant improvement in the detection accuracy. Alternative techniques that could be considered in future include:

Image Combining edge detected imagery in the pre-processing space.
Image Weighting the linear combination according to the degree of information provided by the input image.
Image Performing the fusion process in the feature space by comparing the identified features.

Reducing the false alarm rate.

The application of a set of morphological rules helps to remove a number of obviously un-crater-like features, that otherwise may score highly under the radial consistency transform

Figure 8 shows a false alarm generated using a threshold of ca. 66% with the Radial Consistency algorithm on a region north of the Brent structure. The candidate is approximately 9 km in diameter.

The detected candidate has a certain amount of radial symmetry, but profiles through the elevation map are not suggestive of an impact crater - accordingly the candidate is flagged as a false positive

False

Figure 8. False alarm detected using morphological rules on Landsat, ASAR and SRTM scene.

Similarly, rules based on the textural / spectral characteristics can be added to remove small lakes, reservoirs and other circular features from the detected crater set.


Output   Top

Approximately 20 sites were considered during the course of the study, covering a range of impact crater structures of various sizes and degradation states in both desert and vegetated environments. Two detailed examples are included here

Brent crater, Canada.

The Brent crater is a fairly large simple crater structure which formed about 450 Ma ago. This means that it is a particularly interesting case for crater detection because:

  Image The diameter (~3.8 km) is near the theoretical upper limit for that of a simple Earth impact crater.
  Image The centre of the simple crater has been filled in with about 250 metres of sedimentary rock. The two lakes lie in a pair of hollows in this sedimentary layer. This gives the crater a distinctly different profile from the other simple craters in this section.
  Image The crater rim has been very heavily eroded - and the surrounding vegetation makes it difficult to see the "bowl" shape of the crater at anything other than very low altitudes

Figure 9 shows the results obtained over a scene of 77 km x 85 km using Landsat ETM, ASAR WS and SRTM data using the standard morphological filtering rules and both a 66% threshold on the likelihood score, and the top two scoring features - one of which (towards the centre of the image) is the Brent crater. The image shown is a composite of the input data.

Brent

Figure 9. Brent Crater context, Canada.

Figure 10 shows a 3D projection of the Brent crater using a combination of Landsat ETM band 8 and a resampled SRTM dataset, with a vertical exaggeration of 5. The height is mapped to the red and blue channels to make the topography clearer. The white circle shows the detected crater, demonstrating how the prototype has picked up just inside the crater 'bowl'. In this case the diameter of the detected circle is about 3.2 km.

Brent

Figure 10. Brent Crater 3D projection.

Aorounga, Chad.

Aorounga, in Chad, is a complex structure with a diameter of 12.6 km. The structure consists of a number of circular rings, and has an age of less than 340 Ma. The shape of the crater causes problems for the current rules implemented in the prototype since the morphology is distinctly that of a complex (or even, due to the erosion of the central uplift, small multi-ringed) crater, rather than something that can be modelled as a simple crater - accordingly the prototype algorithms tend to pick up the eroded central uplift rather than the correct diameter.

Figure 11 shows the context of the Aorounga crater. The scene shown is approximately 150 km by 145 km and was processed using Landsat ETM, ASAR WS and SRTM imagery which was then transformed using a MNF (Minimum Noise Fraction) transform in ENVI. This provides an example of an alternative data fusion technique. The impact crater is the circle to the north west of this image, and the remaining false candidates shown are small quasi-circular features with a non-impact morphology. The inset shows a zoomed in area of the image, showing the crater detection.

Aorounga

Figure 11. Aorounga Crater detection.

More examples.

Barringer

Figure 12. Barringer Crater detection.

Tswaing

Figure 13. Tswaing Crater detection.

Gow

Figure 14. Gow Crater detection.

    Top

Contributors to this page: .

Page last modified on Tuesday 14 of December 2010 14:09:26 CET by .