**4. Land mapping applications**

498 Remote Sensing – 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

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

 <sup>2</sup> evaluated true

Along-track direction

> 

(2)

evaluated true

 

(1)

Fig. 5. Example of manual delineation of target area. Target is 5m x 5m, GSD is 0.4m (across-

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

true

*N*

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

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).

true

100% Relative RMSE [%] \* *<sup>N</sup>*

100% 1 Relative bias \*

abundance level considered, or to a combination of both.

track) by 1.0m (along track).

percent:

of the searched target signature abundance in the pixel under test.

future experiments by using thinner strips of material more closely spaced together.

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 chemicals (Rogge et al, 2008).

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

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

Demonstration of Hyperspectral Image Exploitation for Military Applications 501

pictures (R2 = 0.55) with an average difference between the SMA LAI and the ground LAI of less than 0.5 LAI. Figures 8a and 8b show two forest canopy structures, stem density and stand height, as output from the University of Lethbridge Multiple Forward Mode 3-D Canopy Reflectance Model (MFM-3D) applied to the modified geometric optical mutual shadowing model (GOMS) (Peddle et al, 2003). MFM-3D uses a Look-up-tables (LUT) approach based on various ranges and increments of forest structure parameters (density, horizontal & vertical crown radius, crown height and height distribution) as input. The ranges and increments can be determined either from field data or automatically without prior knowledge. Inversion of MFM-3D model produces results when image reflectance values match the modelled reflectance. Field and MFM-3D stand height produced less than 2 m average height difference with the under-estimation of the MFM-3D model attributed to

The addition of above ground restrictions to UNB trafficability model, as determined by the vegetation layers described above, helps produce more refined trafficability classes as illustrated in Figure 9. Figure 10a shows the graphical user interface (GUI) of the route planning tool. Once all the available layers are loaded into the Input Dialog, the user can select from the Interactive Parameters which restrictions to apply for a particular vehicle type. Examples of route planning for four types of military vehicles are shown in Figure 10b with a low environmental concern (not avoiding areas with a high rutting index), and 10c

*R:672nm, G:549nm, B:488nm*

Fig. 6. (a) Sub-image of Probe-1 airborne imagery of CFB Gagetown and (b) landcover classification derived from an evidential reasoning classifier with classes of vegetation and exposed soil and roads. Overall accuracy: 81.8%. Kappa coefficient: 0.78 (D. Peddle, U. of

0 1km

Unclassified Conifer dominant Deciduous dominant Mixed forest Grassland/shrubland Cutblock Exposed soil/road Water

the difficulty in locating neighbouring pixels with similarity to the center pixel.

with a high environmental concern (avoiding areas prone to produce ruts).

**(a)**

**(b)**

Lethbridge).

training. Table 2 provides a summary of the trials and the objectives sought for each application. The primary objective of each trial was the validation of algorithms used to derive vegetation cover information such as type, density and height, the presence of wetlands and the determination of soil type. These surface features are easily derived from hyperspectral imagery and can contribute to improve knowledge of the terrain for the purposes of trafficability and environmental applications. Each trial was conducted in a different vegetation background ranging from various forest biomes (deciduous, mixte, boreal) to prairie grassland. Details regarding each trial, the available ground truth and the algorithms used to analyse the various datasets can be found in (Ardouin et al, 2007).


Table 2. HYMEX land mapping application trials

**CFB Gagetown trial**. One of the objectives of this trial was to improve the trafficability model used by the Army Meteorological Center (AMC) at CFB Gagetown to plan training exercises, avoid erosion by vehicles and promote environmentally sustainable training. The model currently use as input, the soil moisture content simulated by the University of New Brunswick (UNB) Forest Hydrology Model (ForHyM2) which is based on air and ground temperatures, soil type, the amount of precipitation and the wind speed and direction. Improvement of the trafficability model was achieved by the addition of above ground restrictions such as the forest type (hard/softwood), density and height which can be readily derived from hyperspectral remote sensing. Figure 6b shows a vegetation species classification derived from 15m GSD imagery collected by the Probe-1 sensor (Figure 6a). The overall accuracy (81.8%) and Kappa coefficient (0.78) are based on 533 pixels. These results were obtained with the University of Lethbridge Mercury classification algorithm. Figure 7 shows the shadow fraction of the forest canopy which was derived from spectral mixture analysis (SMA) (Peddle & Smith, 2005) along with the sunlit deciduous fraction, the sunlit conifer fraction and the background fraction. The image shadow fraction was found to correlate the best with LAI, as measured on the ground on 29 plots with hemispherical

training. Table 2 provides a summary of the trials and the objectives sought for each application. The primary objective of each trial was the validation of algorithms used to derive vegetation cover information such as type, density and height, the presence of wetlands and the determination of soil type. These surface features are easily derived from hyperspectral imagery and can contribute to improve knowledge of the terrain for the purposes of trafficability and environmental applications. Each trial was conducted in a different vegetation background ranging from various forest biomes (deciduous, mixte, boreal) to prairie grassland. Details regarding each trial, the available ground truth and the

algorithms used to analyse the various datasets can be found in (Ardouin et al, 2007).

CFB Gagetown,

Wainwright, AB

Montmorency Experimental Forest, QC

Jun 2004 Feb 2007 AISA

Table 2. HYMEX land mapping application trials

NB

CFB

**Trial location Date Sensor GSD Background Objectives Application** 

algorithms validation (mixed deciduous) & CFB Gagetown trafficability model

algorithms validation (single deciduous

Forest parameter algorithms validation (mixed conifers), wetland mapping & summer/winter dataset investigation

burnt areas, soil disturbance

species)

trafficability & sustainable training

trafficability

trafficability

sustainable training

Sep 2005 Probe-1 15m Acadian forest Forest parameter

Sep 2006 AISA 4m Boreal/grassland Forest parameter

Boreal/summer Boreal/winter

**CFB Gagetown trial**. One of the objectives of this trial was to improve the trafficability model used by the Army Meteorological Center (AMC) at CFB Gagetown to plan training exercises, avoid erosion by vehicles and promote environmentally sustainable training. The model currently use as input, the soil moisture content simulated by the University of New Brunswick (UNB) Forest Hydrology Model (ForHyM2) which is based on air and ground temperatures, soil type, the amount of precipitation and the wind speed and direction. Improvement of the trafficability model was achieved by the addition of above ground restrictions such as the forest type (hard/softwood), density and height which can be readily derived from hyperspectral remote sensing. Figure 6b shows a vegetation species classification derived from 15m GSD imagery collected by the Probe-1 sensor (Figure 6a). The overall accuracy (81.8%) and Kappa coefficient (0.78) are based on 533 pixels. These results were obtained with the University of Lethbridge Mercury classification algorithm. Figure 7 shows the shadow fraction of the forest canopy which was derived from spectral mixture analysis (SMA) (Peddle & Smith, 2005) along with the sunlit deciduous fraction, the sunlit conifer fraction and the background fraction. The image shadow fraction was found to correlate the best with LAI, as measured on the ground on 29 plots with hemispherical

4m 4m

CFB Suffield, AB Sep 2006 AISA 4m Prairie grassland Map invasive species,

pictures (R2 = 0.55) with an average difference between the SMA LAI and the ground LAI of less than 0.5 LAI. Figures 8a and 8b show two forest canopy structures, stem density and stand height, as output from the University of Lethbridge Multiple Forward Mode 3-D Canopy Reflectance Model (MFM-3D) applied to the modified geometric optical mutual shadowing model (GOMS) (Peddle et al, 2003). MFM-3D uses a Look-up-tables (LUT) approach based on various ranges and increments of forest structure parameters (density, horizontal & vertical crown radius, crown height and height distribution) as input. The ranges and increments can be determined either from field data or automatically without prior knowledge. Inversion of MFM-3D model produces results when image reflectance values match the modelled reflectance. Field and MFM-3D stand height produced less than 2 m average height difference with the under-estimation of the MFM-3D model attributed to the difficulty in locating neighbouring pixels with similarity to the center pixel.

The addition of above ground restrictions to UNB trafficability model, as determined by the vegetation layers described above, helps produce more refined trafficability classes as illustrated in Figure 9. Figure 10a shows the graphical user interface (GUI) of the route planning tool. Once all the available layers are loaded into the Input Dialog, the user can select from the Interactive Parameters which restrictions to apply for a particular vehicle type. Examples of route planning for four types of military vehicles are shown in Figure 10b with a low environmental concern (not avoiding areas with a high rutting index), and 10c with a high environmental concern (avoiding areas prone to produce ruts).

Fig. 6. (a) Sub-image of Probe-1 airborne imagery of CFB Gagetown and (b) landcover classification derived from an evidential reasoning classifier with classes of vegetation and exposed soil and roads. Overall accuracy: 81.8%. Kappa coefficient: 0.78 (D. Peddle, U. of Lethbridge).

Demonstration of Hyperspectral Image Exploitation for Military Applications 503

**Current model Improved model**

Poor (too wet)

Moderate

Soil moisture

Fig. 9. Off-road trafficability classes for the current (left) and improved models (right).

Highly restricted

good Moderate Highly

Excellent Moderate risky

**(b)**

**(c)**

Fig. 10. (a) Route planning for four military vehicles with (b) low and (c) high environmental

**CFB Wainwright trial.** A similar trafficability model was also constructed by UNB for CFB Wainwright with the objective to demonstrate its portability to different soil and vegetation ecosystems. The above ground trafficability is determined by a semi-arid prairie landscape with well defined dry and wet seasons. The vegetation is composed of grassland with areas of deciduous trees (aspen, balsam poplar and willow) and the topography is gentle. The hyperspectral vegetation products made available to UNB are the land cover classes and the leaf area index (LAI) (Figures 11b and 11c) provided by York University under contract to HYMEX. Figures 11d to 11f show the optimal route planning between point A and point B using a Wolf and a LAV vehicles when constrained by wet areas and vegetation during the wet season (Figure 11d), when constrained by wet areas and vegetation during the dry season (Figure 11e) and when constrained by wet areas and vegetation during the wet season and the requirement to move along tree lines as closely as possible (Figure 11f).

Vegetation density

No Go zone

restricted

Poor (too dense)

0 1 km

0 1 km

Poor

Caution

Soil moisture

**(a)**

concern.

Good mobility

Avoid

Fig. 7. Sub-image of Probe-1 airborne imagery of CFB Gagetown showing the shadow fraction of the forest canopy derived from spectral mixture analysis (D. Peddle, U. of Lethbridge).

Fig. 8. Sub-image of Probe-1 airborne imagery of CFB Gagetown showing forest structural parameters (a) stand density and (b) stand height as derived from the MFM-3D model (D. Peddle, U. of Lethbridge).

Fig. 7. Sub-image of Probe-1 airborne imagery of CFB Gagetown showing the shadow fraction of the forest canopy derived from spectral mixture analysis (D. Peddle, U. of

Fig. 8. Sub-image of Probe-1 airborne imagery of CFB Gagetown showing forest structural parameters (a) stand density and (b) stand height as derived from the MFM-3D model (D.

Lethbridge).

**a)**

**(b)**

Peddle, U. of Lethbridge).

0

0 m

6.5

0 Stem/ha

5000

3

Fig. 10. (a) Route planning for four military vehicles with (b) low and (c) high environmental concern.

**CFB Wainwright trial.** A similar trafficability model was also constructed by UNB for CFB Wainwright with the objective to demonstrate its portability to different soil and vegetation ecosystems. The above ground trafficability is determined by a semi-arid prairie landscape with well defined dry and wet seasons. The vegetation is composed of grassland with areas of deciduous trees (aspen, balsam poplar and willow) and the topography is gentle. The hyperspectral vegetation products made available to UNB are the land cover classes and the leaf area index (LAI) (Figures 11b and 11c) provided by York University under contract to HYMEX. Figures 11d to 11f show the optimal route planning between point A and point B using a Wolf and a LAV vehicles when constrained by wet areas and vegetation during the wet season (Figure 11d), when constrained by wet areas and vegetation during the dry season (Figure 11e) and when constrained by wet areas and vegetation during the wet season and the requirement to move along tree lines as closely as possible (Figure 11f).

Demonstration of Hyperspectral Image Exploitation for Military Applications 505

Tamarack trees thrive in open areas because of their intolerance to shade and their resentment to compete with other species (Beeftink, 1951). They also adapt very well to poorly drained soil. Their presence is generally associated with peatlands although their absence do not indicate that there is no wetlands. Spectrally, tamaracks are similar to other conifers in summer and to deciduous trees in winter because they loose their leaves (needles) before winter. A combination of summer and winter data allows the exploitation of this unique characteristic of the tamarack trees to locate and map treed wetlands. Figure 12b shows a RGB of AISA data acquired in winter 2007 at the MEF. Figure 12c is a moisture stress vegetation index in which the red color represents exposed bark. When looking closer at the single tree where the red arrow is pointing in Figure 12a, it is easy to recognise from the shape of its shadow on the snow that it is a defoliated coniferous tree. Not having this information on hand it would be difficult to determine whether the trees in the red color class

Under contract to DRDC, Laval University applied four filters, each made of a band ratio index and a predefined threshold, to classify the tamarack trees in the AISA image with 95% of the tamarack pixels correctly classified and only 1.2% of the remaining pixels misclassified as hardwood trees. Each filter discriminates tamarack trees from other forest features such as other softwood and deciduous. The effects of the application of the first 3

(a) (b) (c)

Fig. 12. (a) high resolution color image, (b) AISA color composite (R:1290 nm , G:1655 nm, B:2189 nm), and (c) a moisture stress vegetation index from the AISA winter image of an area at the Montmorency experimental forest. The red arrow points at a tamarack tree.

**CFB Suffield trial.** The objectives of this trial were primarily environmentally oriented. Despite a semi-arid climate, prairie grasslands are very sensitive to the introduction of invasive species which are often dispersed during military training and along the maintenance roads of pipelines and gas wells. Leafy spurge and crested wheat are the main invasive species and can easily spread in windward direction into preserved native prairie areas. Moreover, there is a need for monitoring training areas for an environmentally sustainable training. This is to ensure that excessive training does not over stress the soil and

in Figure 12c are deciduous, dead spruce or fir, or dormant tamarack.

and the first 4 filters are shown in Figure 13a and 13b.

Fig. 11. (a) AISA color composite, (b) land cover classes and (c) leaf area index map of CFB Wainwright (York University). Route between A and B for a Wolf and a LAV vehicles: (d) when constrained by wet season conditions, (e) when constrained by dry season conditions and (f) when constrained by wet season conditions and the requirement to move along tree lines (UNB).

**Montmorency Experiment Forest (MEF) trials.** Two airborne hyperspectral datasets are available for the MEF site, one from Jun 2004 (summer) and one from Feb 2007 (winter). The objective for imaging this site was to test algorithms for vegetation mapping in a coniferous dominant forest ecosystem. The winter 2007 dataset was acquired for the purpose of investigating the usefulness of summer/winter data to extract relevant terrain information for trafficability in the boreal forest. One first attempt to address this later objective is to map tamarack trees (*Larix laricina*) which can be used as an indicator species for the location of peatlands areas dominated by trees, a wetland type being of interest for trafficability.

Road 0 5

0 1km

R: 800nm G: 680nm B: 550 nm

lines (UNB).

Grass Shrubs Low-density trees High-density trees

(a) (b) (c)

(d) (e) (f)

Fig. 11. (a) AISA color composite, (b) land cover classes and (c) leaf area index map of CFB Wainwright (York University). Route between A and B for a Wolf and a LAV vehicles: (d) when constrained by wet season conditions, (e) when constrained by dry season conditions and (f) when constrained by wet season conditions and the requirement to move along tree

**Montmorency Experiment Forest (MEF) trials.** Two airborne hyperspectral datasets are available for the MEF site, one from Jun 2004 (summer) and one from Feb 2007 (winter). The objective for imaging this site was to test algorithms for vegetation mapping in a coniferous dominant forest ecosystem. The winter 2007 dataset was acquired for the purpose of investigating the usefulness of summer/winter data to extract relevant terrain information for trafficability in the boreal forest. One first attempt to address this later objective is to map tamarack trees (*Larix laricina*) which can be used as an indicator species for the location of peatlands areas dominated by trees, a wetland type being of interest for trafficability. Tamarack trees thrive in open areas because of their intolerance to shade and their resentment to compete with other species (Beeftink, 1951). They also adapt very well to poorly drained soil. Their presence is generally associated with peatlands although their absence do not indicate that there is no wetlands. Spectrally, tamaracks are similar to other conifers in summer and to deciduous trees in winter because they loose their leaves (needles) before winter. A combination of summer and winter data allows the exploitation of this unique characteristic of the tamarack trees to locate and map treed wetlands. Figure 12b shows a RGB of AISA data acquired in winter 2007 at the MEF. Figure 12c is a moisture stress vegetation index in which the red color represents exposed bark. When looking closer at the single tree where the red arrow is pointing in Figure 12a, it is easy to recognise from the shape of its shadow on the snow that it is a defoliated coniferous tree. Not having this information on hand it would be difficult to determine whether the trees in the red color class in Figure 12c are deciduous, dead spruce or fir, or dormant tamarack.

Under contract to DRDC, Laval University applied four filters, each made of a band ratio index and a predefined threshold, to classify the tamarack trees in the AISA image with 95% of the tamarack pixels correctly classified and only 1.2% of the remaining pixels misclassified as hardwood trees. Each filter discriminates tamarack trees from other forest features such as other softwood and deciduous. The effects of the application of the first 3 and the first 4 filters are shown in Figure 13a and 13b.

Fig. 12. (a) high resolution color image, (b) AISA color composite (R:1290 nm , G:1655 nm, B:2189 nm), and (c) a moisture stress vegetation index from the AISA winter image of an area at the Montmorency experimental forest. The red arrow points at a tamarack tree.

**CFB Suffield trial.** The objectives of this trial were primarily environmentally oriented. Despite a semi-arid climate, prairie grasslands are very sensitive to the introduction of invasive species which are often dispersed during military training and along the maintenance roads of pipelines and gas wells. Leafy spurge and crested wheat are the main invasive species and can easily spread in windward direction into preserved native prairie areas. Moreover, there is a need for monitoring training areas for an environmentally sustainable training. This is to ensure that excessive training does not over stress the soil and

Demonstration of Hyperspectral Image Exploitation for Military Applications 507

(a) (b) (c) (d)

Fig. 14. (a) AISA imagery RGB color composite, (b) soil clay band depth, (c) classification of soil clay band depth and (d) RGB of the clay band depth (red), the most dominant grass endmember (green) and moss covered soil spectra (blue). (B. Rivard, U. of Alberta)

vegetation endmember in red and the two dominant dry vegetation endmembers in green and blue. The green vegetation (Red color) is located in low land areas which are often located in the vicinity of wetlands. Local cattle grazing is allowed in some area of the military base. The dry vegetation shown in blue represents overgrazed areas which can be compared to impacted areas from training exercise in other area of the military base, thus showing the potential for environmental monitoring for sustainable training. The black area in the northern part represents a recently burnt area where vegetation hasn't started to grow

As indicated in the previous section, the Mercury supervised classification algorithm (Peddle & Ferguson, 2002) performed well during HYMEX field trials and as a result was integrated into the HOST software. In order to evaluate its performance in more details, we compared Mercury to all the supervised classification algorithms offered by the ITT's ENVI 4.8 software (Van Chestein, 2011). Two data sets were used, a 15 m spatial resolution Probe-1 hyperspectral image of CFB Gagetown and a 4 m resolution AISA hyperspectral imagery

of CFB Wainwright. The classes defined for each dataset are listed in Table 3.

**RGB:** R: 860nm, G: 650nm, B: 550nm

back.

**4.2 Classification algorithms comparison** 

Fig. 13. (a) result from applying the first 3 filters in which several hardwood pixels are misclassified as tamarack, and (b) result from applying the first 4 filters (red dots are correctly classified tamarack pixels). (Prof. Sylvie Daniel and Gaël Briant, Laval University).

therefore the vegetation capacity to recover. The imaged areas include a wide range of soil and vegetation species including invasive species, burned areas, cultivation and grazing areas, wetlands, and various levels of disturbances by vehicle pathways. At the time of the airborne hyperspectral survey (Sep 2006), the prairie landscape was dry and with the exception of the low lands and around wetlands, the vegetation exhibited a low photosynthetic activity which resulted in less pigment absorption in the visible and more apparent absorption features in the short wave infrared by other plant cell constituents such as lignin and cellulose. The following results obtained by the University of Alberta (under contract for HYMEX) demonstrate the potential of this dataset for mapping soil and vegetation at CFB Suffield to help the environmentally sound planning of military exercises.

Soil was determined to have a high clay content. Thus, exposed soil was mapped using the spatial distribution of the depth of the clay absorption feature in the vicinity of 2200 nm after removing the vegetation effect using an orthogonal subspace projection and known green and dry vegetation endmember spectra (Figure 14b). Band depth was measured using the continuum removal between 2210nm and 2230 nm. The band depth was classified into four classes (Figure 14c) defined as (1) low clay absorption depth (green) corresponding to natural undisturbed terrain, (2) slightly (yellow) and strongly (blue) disturbed soils areas and (4) high clay absorption (red) which correspond to bare soils, active roads, non-vegetated dry wetlands and burnt areas. In undisturbed grassland areas the soil is covered with dry grass, old grass residue and a layer of moss. When the surface is disturbed, some of the soil becomes exposed and the amount of moss and old residues decreases. Thus, a good indicator of vegetation recovery following exercises would be a dominance of dry grass. Figure 14d shows an RGB of the clay band depth (red), the most dominant grass endmember (green) and an endmember associated with spectra of moss covered soil measured with a field spectrometer (blue). These three classes of endmembers can easily be associated with the following three conditions: (1) permanently disturbed areas such as roads and areas surrounding gas wells (red), (2) recently disturbed areas where the moss and old residues are removed (green) and (3) undisturbed areas covered by moss, old residues and grass (blue).

Invasive species could not be spectrally identified due to the overall dryness of the vegetation cover. The RGB composite of Figure 15b displays the most abundant green

(a) (b)

Fig. 13. (a) result from applying the first 3 filters in which several hardwood pixels are misclassified as tamarack, and (b) result from applying the first 4 filters (red dots are correctly classified tamarack pixels). (Prof. Sylvie Daniel and Gaël Briant, Laval University).

therefore the vegetation capacity to recover. The imaged areas include a wide range of soil and vegetation species including invasive species, burned areas, cultivation and grazing areas, wetlands, and various levels of disturbances by vehicle pathways. At the time of the airborne hyperspectral survey (Sep 2006), the prairie landscape was dry and with the exception of the low lands and around wetlands, the vegetation exhibited a low photosynthetic activity which resulted in less pigment absorption in the visible and more apparent absorption features in the short wave infrared by other plant cell constituents such as lignin and cellulose. The following results obtained by the University of Alberta (under contract for HYMEX) demonstrate the potential of this dataset for mapping soil and vegetation at CFB Suffield to help the environmentally sound planning of military exercises. Soil was determined to have a high clay content. Thus, exposed soil was mapped using the spatial distribution of the depth of the clay absorption feature in the vicinity of 2200 nm after removing the vegetation effect using an orthogonal subspace projection and known green and dry vegetation endmember spectra (Figure 14b). Band depth was measured using the continuum removal between 2210nm and 2230 nm. The band depth was classified into four classes (Figure 14c) defined as (1) low clay absorption depth (green) corresponding to natural undisturbed terrain, (2) slightly (yellow) and strongly (blue) disturbed soils areas and (4) high clay absorption (red) which correspond to bare soils, active roads, non-vegetated dry wetlands and burnt areas. In undisturbed grassland areas the soil is covered with dry grass, old grass residue and a layer of moss. When the surface is disturbed, some of the soil becomes exposed and the amount of moss and old residues decreases. Thus, a good indicator of vegetation recovery following exercises would be a dominance of dry grass. Figure 14d shows an RGB of the clay band depth (red), the most dominant grass endmember (green) and an endmember associated with spectra of moss covered soil measured with a field spectrometer (blue). These three classes of endmembers can easily be associated with the following three conditions: (1) permanently disturbed areas such as roads and areas surrounding gas wells (red), (2) recently disturbed areas where the moss and old residues are removed (green) and (3) undisturbed areas covered by moss, old residues and grass (blue). Invasive species could not be spectrally identified due to the overall dryness of the vegetation cover. The RGB composite of Figure 15b displays the most abundant green

Fig. 14. (a) AISA imagery RGB color composite, (b) soil clay band depth, (c) classification of soil clay band depth and (d) RGB of the clay band depth (red), the most dominant grass endmember (green) and moss covered soil spectra (blue). (B. Rivard, U. of Alberta)

vegetation endmember in red and the two dominant dry vegetation endmembers in green and blue. The green vegetation (Red color) is located in low land areas which are often located in the vicinity of wetlands. Local cattle grazing is allowed in some area of the military base. The dry vegetation shown in blue represents overgrazed areas which can be compared to impacted areas from training exercise in other area of the military base, thus showing the potential for environmental monitoring for sustainable training. The black area in the northern part represents a recently burnt area where vegetation hasn't started to grow back.

#### **4.2 Classification algorithms comparison**

As indicated in the previous section, the Mercury supervised classification algorithm (Peddle & Ferguson, 2002) performed well during HYMEX field trials and as a result was integrated into the HOST software. In order to evaluate its performance in more details, we compared Mercury to all the supervised classification algorithms offered by the ITT's ENVI 4.8 software (Van Chestein, 2011). Two data sets were used, a 15 m spatial resolution Probe-1 hyperspectral image of CFB Gagetown and a 4 m resolution AISA hyperspectral imagery of CFB Wainwright. The classes defined for each dataset are listed in Table 3.

Demonstration of Hyperspectral Image Exploitation for Military Applications 509

algorithms are most consistent in accuracy as the number of bands is changed. This way, by using a classification algorithm with known consistency, the optimal band-set can be

The major finding was that the Mercury algorithm consistently provides very high overall classification accuracy values as illustrated in Table 4. It proves stable and offers the advantage of not requiring that the number of training pixels for each class be at least equal to the number of bands used plus one as is the case with the Maximum Likelihood and Mahalanobis Distance techniques. Mercury's accuracy increased with the number of bands and it offered the highest individual accuracy values in both datasets. Using Mercury on the principal components yielded lower accuracy than with the original dataset. With the Maximum Likelihood algorithm applied to the principal components, results were almost identical to those obtained

**Number of bands → 4 7 20 50 Average Mercury** 82.30% 83.60% 88.20% 91.30% **86.4%** 

**Divergence** 68.70% 66.60% 68.40% 67.60% **67.8% Binary Encoding** 52.00% 51.40% 56.60% 66.70% **56.7%** 

Table 4. Comparison of Mercury and ENVI supervised classification algorithms accuracy. Green is for the algorithm that ranked 1st in classification accuracy and yellow is the

The study also highlighted the fact that class accuracy varies greatly with the choice of bands in most algorithms. Figures ranging from 0% to 100% accuracy were observed in

In summary, Mercury compares very favourably with ITT's offering for global and class accuracy and for all algorithms, one would be well advised to run a few tests as to the

The Canadian Centre for Mine Action Technologies initiated a study in 2003 to investigate the possibility of exploiting advances in genetic engineering and plant biotechnology to design a process by which plants, local to a region of interest, could be genetically modified

some algorithms but Mercury came out with very consistent global figures.

number and choice of bands to ensure optimal feature accuracy.

**Support Vector Machine** 84.40% 83.10% 84.40% 85.60% **84.4% Mahalanobis Distance** 77.60% 75.20% 86.20% 89.50% **82.1% Neural Network** 81.60% 84.20% 88.30% 65.20% **79.8% Maximum Likelihood** 87.60% 73.00% 77.40% 75.20% **78.3% Minimum Distance** 80.30% 76.70% 76.60% 76.50% **77.5% Parallelepiped** 75.60% 74.00% 77.40% 77.00% **76.0% Spectral Angle** 67.20% 67.60% 68.80% 69.00% **68.2%** 

selected quickly after performing a few tests.

**Spectral Information** 

algorithm that ranked second.

**4.3 Chemical effects on vegetation 4.3.1 Plants as chemical detectors** 

with the original data. The following table illustrates the findings.

Fig. 15. (a) RGB "true color" (red: 640nm, green: 550nm, blue: 460nm) AISA image. (b) RGB composite image of the most widely spread green vegetation endmember (Red) and two dominant dry vegetation endmembers (green, blue). (B. Rivard, U. of Alberta).


Table 3. Number of training and test pixels for each class: (a) classes for CFB Gagetown data (Probe-1, 15m GSD) and (b) classes for CFB Wainwright data, (AISA, 4m GSD).

During the comparison, it was found that the tested algorithms behaved differently as the number of bands used in the classification process increases. Some see classification accuracy increase, others prove unaffected by the number of bands while a third group see the accuracy decrease, albeit slightly.

An immediate advantage in using fewer bands is that processing times are shorter, which is very convenient when analyzing large files. It can also be useful to identify which

(a) (b)

Fig. 15. (a) RGB "true color" (red: 640nm, green: 550nm, blue: 460nm) AISA image. (b) RGB composite image of the most widely spread green vegetation endmember (Red) and two

Table 3. Number of training and test pixels for each class: (a) classes for CFB Gagetown data

During the comparison, it was found that the tested algorithms behaved differently as the number of bands used in the classification process increases. Some see classification accuracy increase, others prove unaffected by the number of bands while a third group see

An immediate advantage in using fewer bands is that processing times are shorter, which is very convenient when analyzing large files. It can also be useful to identify which

dominant dry vegetation endmembers (green, blue). (B. Rivard, U. of Alberta).

(a) (b)

the accuracy decrease, albeit slightly.

(Probe-1, 15m GSD) and (b) classes for CFB Wainwright data, (AISA, 4m GSD).

algorithms are most consistent in accuracy as the number of bands is changed. This way, by using a classification algorithm with known consistency, the optimal band-set can be selected quickly after performing a few tests.

The major finding was that the Mercury algorithm consistently provides very high overall classification accuracy values as illustrated in Table 4. It proves stable and offers the advantage of not requiring that the number of training pixels for each class be at least equal to the number of bands used plus one as is the case with the Maximum Likelihood and Mahalanobis Distance techniques. Mercury's accuracy increased with the number of bands and it offered the highest individual accuracy values in both datasets. Using Mercury on the principal components yielded lower accuracy than with the original dataset. With the Maximum Likelihood algorithm applied to the principal components, results were almost identical to those obtained with the original data. The following table illustrates the findings.


Table 4. Comparison of Mercury and ENVI supervised classification algorithms accuracy. Green is for the algorithm that ranked 1st in classification accuracy and yellow is the algorithm that ranked second.

The study also highlighted the fact that class accuracy varies greatly with the choice of bands in most algorithms. Figures ranging from 0% to 100% accuracy were observed in some algorithms but Mercury came out with very consistent global figures.

In summary, Mercury compares very favourably with ITT's offering for global and class accuracy and for all algorithms, one would be well advised to run a few tests as to the number and choice of bands to ensure optimal feature accuracy.

#### **4.3 Chemical effects on vegetation**

#### **4.3.1 Plants as chemical detectors**

The Canadian Centre for Mine Action Technologies initiated a study in 2003 to investigate the possibility of exploiting advances in genetic engineering and plant biotechnology to design a process by which plants, local to a region of interest, could be genetically modified

Demonstration of Hyperspectral Image Exploitation for Military Applications 511

This research team examined the spectral response of individual leaves of three common Canadian plant species (poplar (Populus deltoides, Populus trichocarpa), wheat (Triticum aestivum), canola (Brassica napus)), which were subjected to fumigation with gaseous phase toxic industrial chemicals and chemicals precursor to chemical warfare agents (e.g. ammonia and sulphur dioxide) (TICs). Treatments were designed to allow quantification of the variation in spectra that might be expected due to environmental, developmental, and

The test plants were grown in controlled environment chambers at the University of Alberta, using standardized conditions. Each spectral measurement collected with the ASD® FR spectrometer, as shown in Figure 16, consisted of an average of 10 scans. Multiple scans were taken per leaf location to reduce the effects of noise. For each leaf, three different locations were measured located approximately halfway between the main leaf vein and the leaf edge, precluding overlap of areas measured. The measurements from each leaf were then averaged accounting for spectral variability across the leaf. For smaller leaves (e.g. new

Fig. 16. Basic set-up for spectral measurements. Inset is an image of the ASD® Leaf Clip, the

The study was broken into two phases: 1) to capture the spectral variability of the various leaf growth stages (new to senescing leaves) observed in each of the three plant types; and 2) subjecting the plants to environmental stresses (e.g. drought) and the following five industrially relevant gaseous phase TICs: ammonia (NH3), sulphur dioxide (SO2), hydrogen sulphide (H2S), chlorine (Cl2), and hydrogen cyanide (HCN). The experimental data were analyzed to determine if the various treatments resulted in specific leaf spectral features related to TICs. Figure 17 illustrates typical effects of the chemicals on plants and Figure 18 depicts representative spectra collected, in this case for canola exposed to Cl2 and SO2. Here one can see key absorption features observed in endmember spectra, which were exploited

Observations showed that both environmental stress and TIC treatments induce similar spectral features inherent to plants, which can be related primarily to chlorophyll and water loss. These include pigments in the visible and cellulose, lignin, lipids, starches, and sugars in the short wave infrared. Although no specific spectral features could be tied to individual TICs, an analysis of the data using vegetation indices, which focus on key spectral bands associated with chlorophyll, pigments and water content, showed that the TICs and

stochastic effects on the physiological state of individual plants within each species.

growth) only one or two measurements were possible.

field spectrometer used to collect plant data.

in subsequent analysis.

(GM) to be sensitive to the compounds known to permeate the soil around emplaced landmines. In this case it was envisioned that the plant's genes would also be designed to include a reporting mechanism, signalling the presence of these compounds through a change in the plant's structure, appearance or some other physical characteristic. The Deyholos group at the University of Alberta was funded to conduct the initial study (Deyholos et al, 2006).

At the same time, the United States' Defense Advanced Research Projects Agency (DARPA) inititated the Biological Input Output Systems (BIOS) program. The BIOS program's objective was to produce basic biochemical modules for future use in plant or microbialbased detectors of chemical and biological compounds of strategic interest. Collaboration between the two projects advanced efforts in developing a human-readable biological signalling event (Deyholos et al, 2007; Antunes et al, 2006)

The DARPA-funded team at Colorado State University went on to develop the first generation plant-based sensor capable of detecting 2,4,6-TNT in the low ppt (parts per trillion) range. The Canadian effort made significant progress in the development of a rootto-shoot transducer system and an effective visual reporter system (Deyholos, 2009).

This effort clearly demonstrated that plants' natural responses to chemicals in their environment could be harnessed, exploited and enhanced to provide an *in situ* chemical detection capability of remarkable sensitivity. This observation, amongst others, led to a study to investigate whether it might be possible to detect, through optical means, the naturally occurring effects of exposure to various chemical agents on vegetation, by which *in situ* vegetation may provide a highly sensitive stand-off detection capability to chemical exposures occurring at ground level. These agents cause stress and damage to surrounding vegetation the extent of which is dependent on dosage and time of exposure.

#### **4.3.2 Passive detection**

It is well recognized that reflective hyperspectral imagery (400-2500nm) is well suited to analyze vegetation. Under the Canadian Space Agency HERO program, a feasibility study was conducted (Peddle et al, 2008) to determine whether a space-based system such as HERO can be used to detect toxic industrial chemicals indirectly by detecting the stress that these chemical cause on vegetation. Recognizing that this could have a potential military application, we pursued this project under HYMEX by conducting a laboratory evaluation of the stresses caused by various chemicals.

The aim of this investigation was to provide information that would help quantify the potential of reflective hyperspectral imagery for chemical and biological surveillance, reconnaissance involving plants exposed to Toxic Industrial Chemicals and Materials (known as TICs and TIMs, such as Ammonia, Sulphur Dioxide, Chlorine, Hydrogen Sulphide, Hydrogen Cyanide, Cyanides, Phosgene).

The two objectives of this study were to determine if: 1) vegetation subjected to TICs could be distinguished from background vegetation during varying growth stages (new growth to senescence) and environmental stresses; and, 2) different TICs could be distinguished based on the vegetation spectral response. This work was conducted by teams at the University of Alberta (Rivard et al, 2008).

(GM) to be sensitive to the compounds known to permeate the soil around emplaced landmines. In this case it was envisioned that the plant's genes would also be designed to include a reporting mechanism, signalling the presence of these compounds through a change in the plant's structure, appearance or some other physical characteristic. The Deyholos group at the University of Alberta was funded to conduct the initial study

At the same time, the United States' Defense Advanced Research Projects Agency (DARPA) inititated the Biological Input Output Systems (BIOS) program. The BIOS program's objective was to produce basic biochemical modules for future use in plant or microbialbased detectors of chemical and biological compounds of strategic interest. Collaboration between the two projects advanced efforts in developing a human-readable biological

The DARPA-funded team at Colorado State University went on to develop the first generation plant-based sensor capable of detecting 2,4,6-TNT in the low ppt (parts per trillion) range. The Canadian effort made significant progress in the development of a root-

This effort clearly demonstrated that plants' natural responses to chemicals in their environment could be harnessed, exploited and enhanced to provide an *in situ* chemical detection capability of remarkable sensitivity. This observation, amongst others, led to a study to investigate whether it might be possible to detect, through optical means, the naturally occurring effects of exposure to various chemical agents on vegetation, by which *in situ* vegetation may provide a highly sensitive stand-off detection capability to chemical exposures occurring at ground level. These agents cause stress and damage to surrounding

It is well recognized that reflective hyperspectral imagery (400-2500nm) is well suited to analyze vegetation. Under the Canadian Space Agency HERO program, a feasibility study was conducted (Peddle et al, 2008) to determine whether a space-based system such as HERO can be used to detect toxic industrial chemicals indirectly by detecting the stress that these chemical cause on vegetation. Recognizing that this could have a potential military application, we pursued this project under HYMEX by conducting a laboratory evaluation

The aim of this investigation was to provide information that would help quantify the potential of reflective hyperspectral imagery for chemical and biological surveillance, reconnaissance involving plants exposed to Toxic Industrial Chemicals and Materials (known as TICs and TIMs, such as Ammonia, Sulphur Dioxide, Chlorine, Hydrogen

The two objectives of this study were to determine if: 1) vegetation subjected to TICs could be distinguished from background vegetation during varying growth stages (new growth to senescence) and environmental stresses; and, 2) different TICs could be distinguished based on the vegetation spectral response. This work was conducted by teams at the University of

to-shoot transducer system and an effective visual reporter system (Deyholos, 2009).

vegetation the extent of which is dependent on dosage and time of exposure.

signalling event (Deyholos et al, 2007; Antunes et al, 2006)

(Deyholos et al, 2006).

**4.3.2 Passive detection** 

Alberta (Rivard et al, 2008).

of the stresses caused by various chemicals.

Sulphide, Hydrogen Cyanide, Cyanides, Phosgene).

This research team examined the spectral response of individual leaves of three common Canadian plant species (poplar (Populus deltoides, Populus trichocarpa), wheat (Triticum aestivum), canola (Brassica napus)), which were subjected to fumigation with gaseous phase toxic industrial chemicals and chemicals precursor to chemical warfare agents (e.g. ammonia and sulphur dioxide) (TICs). Treatments were designed to allow quantification of the variation in spectra that might be expected due to environmental, developmental, and stochastic effects on the physiological state of individual plants within each species.

The test plants were grown in controlled environment chambers at the University of Alberta, using standardized conditions. Each spectral measurement collected with the ASD® FR spectrometer, as shown in Figure 16, consisted of an average of 10 scans. Multiple scans were taken per leaf location to reduce the effects of noise. For each leaf, three different locations were measured located approximately halfway between the main leaf vein and the leaf edge, precluding overlap of areas measured. The measurements from each leaf were then averaged accounting for spectral variability across the leaf. For smaller leaves (e.g. new growth) only one or two measurements were possible.

Fig. 16. Basic set-up for spectral measurements. Inset is an image of the ASD® Leaf Clip, the field spectrometer used to collect plant data.

The study was broken into two phases: 1) to capture the spectral variability of the various leaf growth stages (new to senescing leaves) observed in each of the three plant types; and 2) subjecting the plants to environmental stresses (e.g. drought) and the following five industrially relevant gaseous phase TICs: ammonia (NH3), sulphur dioxide (SO2), hydrogen sulphide (H2S), chlorine (Cl2), and hydrogen cyanide (HCN). The experimental data were analyzed to determine if the various treatments resulted in specific leaf spectral features related to TICs. Figure 17 illustrates typical effects of the chemicals on plants and Figure 18 depicts representative spectra collected, in this case for canola exposed to Cl2 and SO2. Here one can see key absorption features observed in endmember spectra, which were exploited in subsequent analysis.

Observations showed that both environmental stress and TIC treatments induce similar spectral features inherent to plants, which can be related primarily to chlorophyll and water loss. These include pigments in the visible and cellulose, lignin, lipids, starches, and sugars in the short wave infrared. Although no specific spectral features could be tied to individual TICs, an analysis of the data using vegetation indices, which focus on key spectral bands associated with chlorophyll, pigments and water content, showed that the TICs and

Demonstration of Hyperspectral Image Exploitation for Military Applications 513

Fig. 19. A selection of vegetation indices across all species for treatments with NaCl, NH3, SO2, Cl2, HCN, and, dehydration (H2O), senescence and controls plants. The existence of

While the exact physiological response to each stressor remains to be understood, the existence of species-specific responses of vegetation to TICs presents both a challenge and

The HYMEX project studied several potential marine applications in collaboration with Borstad Associates and the Dept. of Fisheries and Ocean (Institut Maurice Lamontagne and Bedford Institute of Oceanography). We conducted airborne hyperspectral surveys on East and West coasts of Canada to evaluate algorithms for near-shore bathymetry, beach trafficability, near-shore bottom type mapping as well as retrieval of chlorophyll and suspended matter concentrations as indicator of water clarity. More details are provided in (Ardouin, 2007). Through this work it was realized that most of these algorithms can be applied to multispectral imagery and that their experimental validation is difficult. The later is particularly true for products that vary with time (or current) and thus would require many measurement stations (for validation) that would need to operate coincidentally with

More recently, we tasked OEA Technologies to provide an operational assessment of HYMEX marine algorithms. In this assessment, a distinction was made between dynamic (e.g. water color) and static (e.g. bathymetry) products. It was pointed out that the Canadian Forces needs for off-shore dynamic products (e.g. water colour) is already fulfilled by marine multispectral sensors (MERIS, MODIS) with pixel size > 250m (Williams, 2009). There might however be a niche for hyperspectral sensors (airborne and spaceborne) which

species-specific responses of vegetation to TICs presents both a challenge and an

the airborne survey and be distributed over the area of the survey.

opportunity for regional remote sensing.

an opportunity for regional remote sensing.

**5. Marine algorithms** 

Fig. 17. Examples of environmental (H2O, NaCl) and chemicals (NH3) stresses on plants.

Fig. 18. Representative spectra for canola experiments using Cl2 and SO2. Mean spectra of control G1 (black line, mature healthy) and G4C (pink line, high senescence) are included for reference. Arrows denote key absorption features observed in endmember spectra compared with G1. Black dotted box denotes smoothing of red edge. Chemical exposure differences is apparent.

environmental stresses result in diagnostic light reflectance data trends from healthy mature to highly stressed leaves.

Comparison of relevant vegetation indices, such as that depicted in Figure 19, showed that specific combinations could be used to distinguish NH3, SO2, Cl2 consistently across all three species (Rogge et al, 2008). The trends result from the variable leaf response within plants, between plants and between species and it is expected much of the variability observed within species would be preserved or even enhanced in nature. As such it is encouraging for the possible detection of TIC effects on natural vegetation using airborne/spaceborne imagery.

As the detection methodology was developed from leaf-level observations, it is important to note that field trials remain to be conducted in order to test if the findings of this study can be extended to the detection of TICs in the natural environment. The principal unknown is the effect of varying vegetation canopy structural parameters (e.g. canopy gaps, leaf area) and background properties (litter and soil reflectance) on the specific data trends that were identified.

Fig. 19. A selection of vegetation indices across all species for treatments with NaCl, NH3, SO2, Cl2, HCN, and, dehydration (H2O), senescence and controls plants. The existence of species-specific responses of vegetation to TICs presents both a challenge and an opportunity for regional remote sensing.

While the exact physiological response to each stressor remains to be understood, the existence of species-specific responses of vegetation to TICs presents both a challenge and an opportunity for regional remote sensing.
