**3. Target detection algorithms**

Throughout the HYMEX project, DRDC gained experience in applying algorithms for target detection applications. In this section, we describe a typical processing chain (atmospheric correction, detection and target abundance estimation (Roy, 2010)) used in the project and present results from an experiment aimed at evaluating the performance of the target abundance estimation part of the processing chain using data collected in difficult illumination and atmospheric conditions.

In late October and early November 2009, DRDC collected airborne hyperspectral imagery near Suffield, Alberta (50°13'N, 110°10'W) using an Itres SASI-600 SWIR pushbroom imaging system. The sensor was flown at various altitudes ranging from 330m to 1700m above ground level in order to acquire imagery at across-track ground sampling distance (GSD) of 0.4m, 1.0m, and 2.0m, while along-track GSD remained constant at 1.0m. Imagery was collected between 13h00 and 15h00 local time, which resulted in sun elevation between 17 and 25 degrees. Furthermore, thin altostratus clouds and an overcast of altocumulus clouds on the first (29 Oct) and second day (03 Nov) of collect respectively degraded the illumination conditions considerably, as illustrated in Figure 2. Compared to typical reflective hyperspectral field trials usually conducted under clear skies and high solar elevation, this data collection was conducted under significantly adverse environmental conditions not often considered in the hyperspectral literature.

Fig. 2. Typical sky conditions on 29 Oct 2009 (left) and 03 Nov 2009 (right).

One objective of this field trial was the evaluation of the constrained energy minimisation (CEM) algorithm (Settle, 2002) sub-pixel abundance estimation accuracy. For this purpose, we designed targets of known abundances made of thin strips of painted metal, as illustrated in Figure 3 below. The design allowed changes to the abundance level by varying the distance between the strips of metals while their overall size (5m x 5m) ensured that they filled completely at least one pixel in the imagery, as showed in Figure 4. We used two different types of paint to vary the contrast between the target and background, one beige (see Figure 3) and one green (not shown). The base color was mixed with small quantities (2 to 10% per volume) of black feature-less paint to control the spectral features depth and overall signature albedo. A total of 6 targets were used in this experiment.

processing, atmospheric correction, exploitation and interactive tools. A Navigation Tool also allows loading customized task descriptions with links to the HOST user interface. This guides the user through the algorithms needed to accomplish a task and the selection of the

Throughout the HYMEX project, DRDC gained experience in applying algorithms for target detection applications. In this section, we describe a typical processing chain (atmospheric correction, detection and target abundance estimation (Roy, 2010)) used in the project and present results from an experiment aimed at evaluating the performance of the target abundance estimation part of the processing chain using data collected in difficult

In late October and early November 2009, DRDC collected airborne hyperspectral imagery near Suffield, Alberta (50°13'N, 110°10'W) using an Itres SASI-600 SWIR pushbroom imaging system. The sensor was flown at various altitudes ranging from 330m to 1700m above ground level in order to acquire imagery at across-track ground sampling distance (GSD) of 0.4m, 1.0m, and 2.0m, while along-track GSD remained constant at 1.0m. Imagery was collected between 13h00 and 15h00 local time, which resulted in sun elevation between 17 and 25 degrees. Furthermore, thin altostratus clouds and an overcast of altocumulus clouds on the first (29 Oct) and second day (03 Nov) of collect respectively degraded the illumination conditions considerably, as illustrated in Figure 2. Compared to typical reflective hyperspectral field trials usually conducted under clear skies and high solar elevation, this data collection was conducted under significantly adverse environmental

parameters for those algorithms.

**3. Target detection algorithms** 

illumination and atmospheric conditions.

conditions not often considered in the hyperspectral literature.

Fig. 2. Typical sky conditions on 29 Oct 2009 (left) and 03 Nov 2009 (right).

overall signature albedo. A total of 6 targets were used in this experiment.

One objective of this field trial was the evaluation of the constrained energy minimisation (CEM) algorithm (Settle, 2002) sub-pixel abundance estimation accuracy. For this purpose, we designed targets of known abundances made of thin strips of painted metal, as illustrated in Figure 3 below. The design allowed changes to the abundance level by varying the distance between the strips of metals while their overall size (5m x 5m) ensured that they filled completely at least one pixel in the imagery, as showed in Figure 4. We used two different types of paint to vary the contrast between the target and background, one beige (see Figure 3) and one green (not shown). The base color was mixed with small quantities (2 to 10% per volume) of black feature-less paint to control the spectral features depth and

Fig. 3. Left) Example of controlled abundance target. Right) Spectral signatures of the beige paint at different albedo levels, as measured in field conditions using an ASD FieldSpec Pro spectrometer.

Fig. 4. Left) SWIR 3-colors composite of targets, imaged at a 2.4m GSD (coarsest resolution). Right) Background (red) and target (blue) signatures as measured by the airborne sensor, compared to the target library signature (black).

DRDC favours automated and adaptive approaches to hyperspectral target detection, minimizing user interaction and processing time as much as possible. In this context, we used the following processing chain for this dataset exploitation:


Demonstration of Hyperspectral Image Exploitation for Military Applications 499

The results demonstrate the robustness of the processing chain; with minimal user interaction and using a simple processing chain suitable for near real-time exploitation, targets can be characterized at the sub-pixel level even under adverse illumination conditions. This demonstrates the processing chain's military utility, and indicates that it could be adapted to the detection and characterization of spectral signatures of interest in a

Land mapping applications were studied in collaboration with the University of New-Brunswick, the University of Alberta, the University of Lethbridge, York University and Laval University. The work was oriented towards soil and vegetation characterization and mapping for trafficability and environmental applications. In Section 4.1, algorithms for classification and the extraction of vegetation canopy attributes (density, structure) were evaluated using airborne hyperspectral data acquired over three Canadian Forces bases (CFB). The resulting validated hyperspectral products were then used to improve a trafficability model developed by the University of New Brunswick for Gagetown military base as well as promote environmentally sustainable training on military bases. Winter airborne images were also acquired over the Montmorency experimental forest (near Quebec City) to investigate the potential of winter imagery to better derive forest information. In Section 4.2 we show that among the classification algorithms that were evaluated, the Mercury algorithm (an evidential-reasoning-based supervised classification algorithm developed by the University of Lethbridge (Peddle & Ferguson, 2002)) achieved the best performance. Finally, Section 4.3 shows results from a laboratory study conducted by the University of Alberta demonstrating how hyperspectral techniques can be used to discriminate between vegetation stresses caused by exposure to different toxic industrial

This section presents results obtained for the two main land applications of HYMEX, trafficability and the monitoring of training ranges to promote environmentally sustainable

Target type Target configuration Contrast type Albedo type = 0.5 (29 Oct 2009) = 0.33 (03 Nov 2009)

High (beige) High 11.7 30.0 High (beige) Medium 13.4 24.3 High (beige) Low 12.8 19.6 Low (green) High 11.9 20.8 Low (green) Medium 13.2 24.5 Low (green) Low 13.2 21.7 Average over all targets: 12.7 23.5 Table 1. Relative RMS errors of the retrieved abundances using the CEM algorithm;

denotes the target abundance.

military operational context.

chemicals (Rogge et al, 2008).

**4.1 Trafficability and the monitoring of military training areas** 

**4. Land mapping applications** 

4. Finally, we calculated the score for all pixels of the image using the CEM algorithm. As described in Settle (2002), when properly normalised, the CEM output is an estimation of the searched target signature abundance in the pixel under test.

The controlled ground targets were collected in 45 different images. On 29 Oct, the abundance was set to 50% while on 03 Nov it was reduced to 33%. All images were manually interpreted to delineate the area of the targets in the images. The CEM scores were averaged over each target area to derive an "average abundance", as shown in Figure 5. This was necessary because as imaged, the targets had inhomogeneous abundance over their physical extent, particularly at the finest GSDs. This suggests that the target design could be improved for future experiments by using thinner strips of material more closely spaced together.

Along-track direction

Fig. 5. Example of manual delineation of target area. Target is 5m x 5m, GSD is 0.4m (acrosstrack) by 1.0m (along track).

Estimated abundance error was calculated using the root mean square error (RMSE) and the estimation bias, both normalized by the true abundance in order to get a relative error in percent:

$$\text{Relative RMSE} \left[ \% \right] = \frac{100\%}{\alpha\_{\text{true}}} \sqrt{\frac{\sum \left( a\_{\text{evaluated}} - a\_{\text{true}} \right)^2}{N}} \tag{1}$$

$$\text{Relative bias} = \frac{100\%}{a\_{\text{true}}} \ast \frac{1}{N} \sum (a\_{\text{evaluated}} - a\_{\text{true}}) \tag{2}$$

RMSE results are presented in Table 1. Overall, the root mean square error for this experiment is between 11.7% and 30%. In absolute terms, the overall RMSE translates to 0.064 and 0.078 for the 0.5 and 0.33 abundance targets respectively. The retrieved abundances were slightly underestimated, with bias of -1.9% and -14.7% again on the 0.5 and 0.33 abundance targets respectively. Since atmospheric conditions degraded between the two collects, it is unclear if the observed increase in error is related to the change in illumination conditions, to the lower abundance level considered, or to a combination of both.

The results achieved are encouraging and show that target abundance can be retrieved at the subpixel level using the CEM algorithm with a high accuracy. The fact that the estimated abundances are generally lower than the true abundances which is consistent with an error that could have been introduced during the manual delineation of targets area, by assigning larger areas to targets than their true area. Also, the imaging system true point spread function has not been characterized and taken into account in this analysis; non-uniform sampling over the GSD could lead to an underestimation of the sub-pixel abundances (Settle, 2004).


Table 1. Relative RMS errors of the retrieved abundances using the CEM algorithm; denotes the target abundance.

The results demonstrate the robustness of the processing chain; with minimal user interaction and using a simple processing chain suitable for near real-time exploitation, targets can be characterized at the sub-pixel level even under adverse illumination conditions. This demonstrates the processing chain's military utility, and indicates that it could be adapted to the detection and characterization of spectral signatures of interest in a military operational context.
