International Network of Engineers and Scientists Against Proliferation

Detection of Clandestine Nuclear-Weapons-Useable Materials Production with Satellite Imagery

Gotthard Stein

With the availability of satellite imagery, remote sensing plays an important role in nuclear verification and safeguarding. Due to technical improvements in spatial and spectral resolution, satellite imagery can now build the basis of complex systems for recognizing and monitoring even small-scale and short-term structural features of interest within nuclear facilities. The increasing amount of data demands new methodologies and higher efforts regarding image preprocessing, analysis and interpretation. Although a software system cannot totally replace an image analyst, there exist many possibilities to obtain faster and more precise image analysis results, e.g., a pre-selection of objects or the automatic detection and classification of changes. This paper proposes an automated object-based analysis and change detection approach for remote sensing data. New methodologies will be presented for (pre-) processing, analysis, change detection and interpretation as well as their exemplary application on nuclear facilities in Iran. This paper focuses on the use of high-resolution satellite data to detect buildings and other man-made objects since their changes over time may indicate a violation of the Non-Proliferation Treaty. Lower resolution data is used for wide-area monitoring in order to detect undeclared nuclear industrial sites. With the methodology presented here it is possible to identify both industrial sites as well as man-made objects in heterogeneous high-resolution images.

In 2005, at the 49th Annual Regular Session of the International Atomic Energy Agency’s (IAEA) General Conference, the IAEA Director General, Dr. ElBaradei, pointed out, that providing assurance about Iran’s nuclear programme is a great challenge of utmost importance.

With the availability of high-resolution satellite imagery, remote sensing has played an important role in nuclear verification and safeguarding. In light of expected technical improvements in spatial and spectral resolution, satellite imagery can build the basis of future complex systems for recognizing and monitoring even small-scale and shortterm structural features of interests within nuclear facilities. Examples include the construction of new buildings, plant expansions, changes in the operational status and preparations for underground activities.

Large volumes of satellite data demand highly automated image processing, analysis and interpretation. Although image analysts can hardly be replaced by a software system, the potential to obtain faster and more precise image analysis results is considerable. Among the options under consideration are the pre-selection of relevant objects or the automatic detection and classification of changes.

This paper proposes a new methodology for the evaluation of satellite imagery which can cope with this challenge. Iranian facilities are used to illustrate the method. In the context of nuclear safeguards-related monitoring, aspects of automation, standardization and transferability will be discussed.

Preprocessing of Image Data

Differences in radiance values indicating significant (‘real’) changes must be larger than radiance changes due to other factors. The aim of pre-processing is, therefore, to correct the radiance differences caused by other factors such as variations in solar illumination, atmospheric conditions, sensor performance and geometric distortion.

Precise geometric correction is essential to exact pixel-by-pixel or object- by-object comparisons in attempting to detect change. By means of geometric correction algorithms, image data can be cross-registered (image-toimage registration) or registered to a given map projection (geo-referencing). In the presented test cases, a semiautomated image-to-image registration is carried out based on image correlation algorithms with sub-pixel accuracy (RMS error well below +/-1 pixel).[1]

Radiometric correction, which is necessary to obtain absolute surface radiance or reflectance, is achieved by removing atmospheric effects. Absolute atmospheric modelling is rarely required to ensure transferability of image classification models or change detection applications of satellite imagery. Relative radiometric normalization seems to be sufficient, assuming changes in at-sensor radiances can be linearly approximated. Here, relative radiometric normalization based on the so-called no-change pixels is applied to the image data.[2]

The original QuickBird image data used for this case study contain a lower-resolution multispectral data set (2.4 m) and a high-resolution panchromatic image (0.6 m). To obtain a highresolution multispectral data set with 0.6 m ground resolution, a panchromatic sharpening procedure is applied to all data sets. Here, a wavelet-based algorithm is used which produces well-focused results without significantly affecting the original multispectral values.

Object-Based Image Analysis

When applied to high-resolution imagery, traditional pixel-based image processing algorithms often yield limited results. Especially when small structural objects are to be detected, object-based procedures are more appropriate. In comparison to solely spectral features applied within pixelbased approaches, utilization of object features, such as their size, orientation, shape, texture and the relations between the objects in different scales considerably extends the possibilities of image analysis. Analyzing satellite image data with object-based methods also offers the option of refining the image classification or recognition process with more specific knowledge.

Figure 1 illustrates the advantage of a wide feature basis within an object-based classification. The image shows a section of an Iranian nuclear power plant. When only spectral values are considered, it is very difficult to detect the heterogeneous building in the upper left part of the image and the function of the building is almost impossible to determine. The use of objects, on the other hand, provides a very wide feature base. Accordingly, a classification rule for this building could be defined as follows: all objects within the image which are rectangular and have a specific size (shape features) are within a given distance to the reactor dome (position feature) and have smaller rectangular objects on their roofs (objectrelation feature) are generator halls.

The extraction of the objects from the analyzed pre-processed images takes place at the lowest level by segmentation, at which stage the primary segments should ideally represent the real world objects. Feature recognition is carried out with the SEparability and THreshold (SEaTH) analysis tool[3] and provides the basis for image classification.

A pixel-based change detection procedure is applied to the preprocessed data, i.e., to the classified image. With regard to nuclear monitoring, the most satisfactory results were obtained by applying the Multivariate Alteration Detection (MAD) method,[4] which is based on a statistical transformation referred to as canonical correlation analysis.[5] It is used to highlight differences between two scenes which are to be compared by helpint to detect changes in multi-temporal data. For the actual identification of changes, object- oriented image classification and change detection were combined.

Pre-Scanning: Wide-Area Monitoring with Medium-Resolution Satellite Imagery

Due to the potential of dual-use activities associated with the Iranian nuclear program, a nuclear monitoring system was set up for Iran. In the first instance, multitemporal area-wide ASTER imagery (AST\_07, surface reflectance with 15 m (VNIR) and 30 m (SWIR) spatial resolution) for 17 nuclear- related locations built the database for the system, complemented by open source information and, for some areas of interest, high-resolution imagery from QuickBird.

Different image analysis approaches suitable for nuclear safeguards applications were implemented and evaluated. In general, a two-step approach was realised with areas of interest being pre-scanned on the basis of wide-area monitoring with mediumresolution ASTER data. The data could then be further analysed with high-resolution QuickBird image data serving as the basis for change detection and analysis.

The pre-scanning is intended for the detection of potential undeclared nuclear-related activities and of major changes within declared nuclear sites and their surrounding areas. ASTER imagery of the sites located at Arak, Bandar Abbas, Bushehr, Esfahan and Natanz were used as training data in order to determine a fixed set of segmentation parameters for adequate multiresolution object extraction, define satisfactory and transferable object features for object classes relevant to nuclear safeguards and implement a measure for possible changes within nuclear facilities. Figure 2 shows the ASTER data for the sites Bandar Abbas, Bushehr and Natanz.

For the subsequent object classification, a standardised and transferable semantic model was developed which showed satisfactory results for the ASTER (AST\_07) images. In the given project, the optimal object features and the range of their membership functions were automatically determined with SEaTH. In order to detect possible nuclear industrial sites, a classification model was defined for the object classes “industrial sites” and “background” and applied to a number of images. Figure 3 shows the classification result for Bandar Abbas, Bushehr and Natanz. The identified industrial sites have white outlines to improve visibility.

All relevant sites were correctly classified as “industrial sites”. Although some sites were incorrectly marked, there was no need to adapt the classification model for each image, and the results were sufficient to quickly get an overview of the different land cover classes and to detect industrial sites in a wide area. The rule bases that were developed for the object classes can also be applied to the image data individually.

Change Detection and Analysis With High-Resolution Imagery

When areas with significant safeguardsrelated changes have been detected, e.g. on the basis of medium-resolution image data, they are then analyzed in detail with high-resolution imagery.

In the case study, investigations were conducted as to the standardization and transferability of classification models. Esfahan was used as a model.

Esfahan is believed to be one of the centres of the Iranian nuclear program. The Nuclear Fuel Research and Production Center (NFRPC), established in 1974 southeast of the city of Esfahan, is Iran’s biggest nuclear research centre. The NFRPC contains a Miniature Neutron Source Reactor (MNSR), a Light Water Sub-Critical Reactor (LWSCR), a Heavy Water Zero Power Reactor (HWZPR), a Graphite Sub-Critical Reactor (GSCR), a Fuel Fabrication Laboratory (FFL), a Uranium Chemistry Laboratory (UCL), a Uranium Conversion Facility (UCF) as well as a Fuel Manufacturing Plant (FMP). Some of these are still under construction or have already been shut down.

Aster satellite data at 15 m ground resolution was used for the wide area site monitoring. For observation of individual facilities over time, high spatial resolution QuickBird images at 0.6 m ground resolution were used. The case study uses QuickBird images of the NFRPC Esfahan taken in July 2002 and July 2003. This paper cannot give details of the whole monitoring procedure, including preprocessing, modelling, classification and change detection. Rather, it focuses on the results of automated pre-processing, object- based classification and automated change detection of the two pictures.

Pre-processing of the satellite images was done as described above. Figure 4 shows the QuickBird images of the NFRPC Esfahan taken in 2002 and 2003, respectively, after preprocessing. Objects were extracted by means of a multiscale segmentation with eCognition, an image analysis software. This lead to a hierarchy of image objects, the individual features of the Esfahan site being identified at the lowest segmentation level. By increasing the scale parameter, the objects become coarser, until all structures of interest were included in one object.

The multi-resolution segmentation algorithm of eCognition was used with standardized parameters. Feature extraction and semantic modelling with SEaTH was performed for each of the defined object classes “Built-up Areas” and “Streets.” This lead to a ruled-based, object-oriented classification model for the Esfahan NFRPC shown in Figure 5 (left). The overall classification accuracy for the 2002 image was 91%.

A combination of pixel-based change detection and object-based image classification leads to the detection and identification of significant changes in multitemporal data. Figure 5(right) shows a result of this change detection methodology. In this zoom to a subset of the image, new built-up areas and streets can be clearly identified.

Conclusions

An object-based analysis methodology for nuclear safeguards purposes was proposed in order to identify objects and detect significant changes within nuclear facilities. With the techniques described in this paper, preprocessing of high resolution data could be automated. Identification of the objects within the nuclear facilities is achieved by object-based image analysis which includes standardized segmentation, statistical feature analysis and rulebased classification. In sum, the classification result for the test case is very satisfying; in particular the buildings are identified with high accuracy.

For advanced analysis of nuclear sites (using high-resolution imagery), a detailed classification model has to be able to differentiate between nuclear and non-nuclear industrial sites and preferably between the different types of facilities within the nuclear sites class, as well. Although the preliminary results of this test case and previous attempts with automated object-based classification on German nuclear power plants have lead to promising results, many more case studies must be performed to gain a comprehensive understanding of the nuclear site signatures that are identifiable in satellite imagery. Furthermore, efforts to automatically extract object features have to be continued, and the classification accuracy in terms of spatial and temporal transferability needs to be assessed in detail.

Satellite imagery will never provide all relevant information needed for nuclear safeguards and security, but represents a crucial source of information. The developments in sensor technologies (spatial and spectral improvements) and thus the increasing application possibilities of satellite imagery for nuclear safeguards have to be continually investigated and evaluated.

This article is based on the work of Sven Nußbaum, Research Center Juelich, already published in: S. Nussbaum, I. Niemeyer and M.J. Canty, Automated Object-oriented Analysis of High Resolution Remote Sensing Data in the Context of NPT Verification Exemplified for Iranian Nuclear Sites, in: Proc. of the 47th INMM Annual Meeting, Nashville, Tennessee, 16-20 July 2006.

---

1.   M. Lehner, Triple stereoscopic imagery simulation and digital image correlation for MEOSS project, Proc. of the ISPRS Commission I Symposium, Stuttgart, 1986, p. 477–484.
2.   M.J. Canty, A.A. Nielsen and M. Schmidt, Automatic radiometric normalization of multispectral imagery, Remote Sensing of Environment 91, 2004, p. 441–451.
3.   S. Nussbaum, I. Niemeyer and M. Canty, Feature recognition in the context of automated object-oriented analysis of remote sensing data monitoring the Iranian nuclear sites, Proc. of the SPIE Europe Symposium on Optics/Photonics in Security & Defence, 26-28 September 2005, Electro-Optical Remote Sensing Vol. 5988, Bruges, Belgium, 2005.
4.   I. Niemeyer, S. Nussbaum, and M. Canty, Automation of change detection procedures for nuclear safeguards-related monitoring purposes, Proc. of the IEEE International Geoscience and Remote Sensing Symposium, IGARSS’05, Seoul, 25-29 July 2005, 2005.
5.   A. Nielsen, K. Conradsen, and J.J. Simpson, Multivariate alteration detection (MAD) and MAF processing in multispectral, bitemporal image data: New approaches to change detection studies, Remote Sensing of Environment 64, pp. 1–19, 1998.