Classification and Identification

AUG Signals has developed unique target classification and identification technologies. Used in conjunction with our innovative detection algorithms, these have been applied to a variety of applications, from vessel and military targets identification to waterborne contaminants identification and mineral/oil signature identification.

Markov chain based classification technique is an expansion of Markov chain CFAR detector. In other words, the Markov chain CFAR detection technique is adapted for the purposes of classification and identification by constructing Markov chain models for all targets, and calculating their corresponding transition matrices. Similar to its detection applications, introducing the new correlation information to target modeling highly improves target classification accuracy.
The Optimal Projection Adaptive Matched Subspace Classification and Identification employs multiple hyperspectral detectors working with decision fusion and Dampster-Shafer algorithms to produce classification and identification results.

An example of hyperspectral military targets identification is provided below.

Figure 1 is the 6th band image generated using hyperspectral sensor Probe-1. There are eight different military targets in the scene. The ground truth is provided in Figure 2 where targets are marked by different colors and symbols.

classification1

Figure 1. The 6th band image generated using hyperspectral sensor Probe-1.
Figure 2. Ground truth: Red: T1; Green: T2; Yellow: T3; Cyan: T4; Magenta: T5; Black: T6; Blue: T7; Red star: T8; White: Background

The state-of-the-art spectral analysis technology, as well as AUG Signals technology are applied to identify these targets. The results are provided in Figures 3 and 4, respectively.

classification3
Figure 3. Identification result using state-of-the-art spectral analysis technology.

classification4
Figure 4. Identification result using technology developed by AUG Signals.

As can be seen in Figure 3, there are many non-target pixels miss-detected as targets. In addition, most of the detected targets are miss-identified into wrong categories of targets. The result generated using our technology is much clearer. All the targets are correctly identified with only few false alarms. In this example, the performance using our technology is 10 times more accurate than the state-of-the-art technology.

In addition to military applications, AUG’s hyperspectral detection and identification technologies are used to generate state-of-the-art mineral maps. AUG Signals’ algorithms allow for unmatched mapping speed and accuracy; mapping of vast areas that used to take up to 5 years to complete can now be accomplished in a matter of hours, with the end-result being significantly more accurate (resolution < 1m2 vs. 2km2).

The following is an example of mineral maps obtained using AUG Signals’ spectral signal processing techniques. The image cube is Probe-1 data obtained in southern Baffin Island. Some preprocessing of the data have been performed (preprocessing includes calibration, atmospheric correction, water and snow masking, etc). Figure 5 is a spatial subset of the original RGB image (Gaussain enhancement applied) of Probe-1 data obtained in Buffin Island. Spectral signatures of seven materials are plotted in Figure 6. Those seven materials of interest are presented in the image scene in a mixed form. AUG Signals’ confidence images of the first three materials are presented in Figure 7 as RGB images. Figure 8 is a classified image of all 7 materials where each color represents a classified material.

h_surveying1

Figure 5. Original RGB color image (Gaussian enhancement applied): R: band 16, G: band 7, B: band 3.
Figure 6. Spectral signatures of 7 materials (derived from data)
 Blue – veg/metagabbro;
 Red – quarzites;
 Green – psammites;
 Dark cyan – metatonalites;
 Magenta – monzo – granites;
 Light cyan – metatonalites;
 Yellow – psammites.

hy_surveying2

Figure 7. AUG Signals’ Confidence image of the first three materials: Red – quarzites; Green – psammites; Blue – veg/metagabbro. (No histogram enhancement)
Figure 8. Classified image of all 7 materials
 Blue – veg/metagabbro;
 Red – quarzites;
 Green – psammites;
 Dark cyan – metatonalites;
 Magenta – monzo – granites;
 Light cyan – metatonalites;
 Yellow – psammites.
Feature-Based Classification and Identification technique entails selecting good features and simultaneously fusing different classifiers. The algorithm has two steps: 1) feature selection, and 2) classifier selection and fusion. This technique employs an effective lower bound estimate for the Bayes Risk and uses it to judge the effectiveness of various classification approaches and features sets. This Bayes Risk estimate is an essential tool for choosing sets of features, since those sets that show lower Bayes Error estimate will have the most potential for developing a superior classification strategy. At the final stage, dramatically improved classification accuracy is guaranteed as a result of fusion of classifiers.

The feature-based classifier has been successfully applied to Radarsat-2 data (with 3 meters resolution) with very high classification accuracy (approximately 0.96). In the testing data, 4 vessel classes were present: CFAV Quest, CCGS Sir Wilfred Grenfell, CCGC Sambro, and Divecom III, with dimensions of 76m x 12.6m, 64.48m x 5.18m, 16.25m x 5.18m, and 13.3m x 4.45m, respectively.

Vessel surfaces contain many spatially distributed scattering centers. After applying different polarimetric decompositions and transformations, the scattering types and locations can be determined, thus differentiating different vessels. For example, the scattering entropy – scattering type angle – plane using Cloude-Pottier decomposition and the unit disc representation of symmetric scattering matrix space using Cameron decomposition have been widely used for classification.

AUG Signals has developed a robust vessel classification system based on polarimetric decomposition and transformation by integrating the company’s advanced CFAR detection algorithm into the polarimetric decomposition classification method, thereby increasing the classification and identification accuracy.

To utilize polarimetric-based vessel classifier, data should have full polarizations to apply polarimetric decomposition and transformation.