Highresolution satellite imagery for detection of ... - Semantic Scholar

3 downloads 41086 Views 1MB Size Report
small tracks and minor damage on the vegetation caused by small ATVs were easily detected. The total length ... it will take years to recover. ... Quickbird which can provide data with a spatial resolution up ..... fen and mire vegetation is commonly used to drive on by. ATVs. .... These areas are very large and therefore hard to.
land degradation & development Land Degrad. Develop. (2010) Published online in Wiley Online Library (wileyonlinelibrary.com) DOI: 10.1002/ldr.1047

HIGH-RESOLUTION SATELLITE IMAGERY FOR DETECTION OF TRACKS AND VEGETATION DAMAGE CAUSED BY ALL-TERRAIN VEHICLES (ATVS) IN NORTHERN NORWAY H. TØMMERVIK1*, B. JOHANSEN2, K.A. HØGDA2 AND K.B. STRANN1 1

Norwegian Institute for Nature Research (NINA), N-9296 Tromsø, Norway 2 NORUT Information Technology, N-9291 Tromsø, Norway

Received 15 June 2010; Revised 12 July 2010; Accepted 27 August 2010

ABSTRACT Use of all-terrain vehicles (ATVs) in vulnerable environments may cause vegetation and soil erosion that will last for a long time. In northern parts of Norway, the growing use of ATVs off-road has made considerable damage to the environment. In Troms County, the Norwegian Army has exercise and shooting ranges (battlefields) where there has been extensive use of light, medium and heavy armored military vehicles and ATVs for decades. The main objective of this study has been to test the feasibility of high spatial resolution satellite imagery as Ikonos and Quickbird in combination with fieldwork to detect tracks and damage caused by heavy armoured military vehicles (tanks) as well as medium and light ATVs. Applying simple image processing techniques, this study showed that larger parts of these areas were influenced and even small tracks and minor damage on the vegetation caused by small ATVs were easily detected. The total length of the tracks within the battlefields was calculated to be c. 1000 km. Using a zone of damage/influence of 50 m on each side of the tracks, the damaged and/or influenced area by ATVs was estimated to cover 56 km2 (17 per cent) of the total area of 334 km2. The conclusion of this study is that highresolution optical satellite images are well suited for surveying damage on the vegetation caused by terrain vehicles in northern Norway. Copyright # 2010 John Wiley & Sons, Ltd. key words: fens; light ATVs; military battlefields; satellite remote sensing; sub-Arctic; Norway; tanks; vulnerable environment

INTRODUCTION The Norwegian Army has a long history for utilizing land and landscape in Norway for shooting and exercises. In Troms County, Northern Norway, the Norwegian army has several battlefields (Figure 1) where there have been extensive use of all-terrain vehicles (ATVs); light ATVs, medium ATVs and heavy armoured vehicles (Figure 2). All these types of vehicles have through decades caused damage to different vegetation types, to land cover, soils, geomorphological features and landscape. Through a growing awareness of land-use damage and an intention of a more sustainable use of these battlefields, the army initiated a project were the long-term effects on the environmental conditions caused by military vehicles and exercises was studied (Tømmervik et al., 2005). Several studies performed both in Norway and Sweden shows that especially the mountain region is vulnerable for use of terrain vehicles (Renman, 1989; Norberg et al., 1998). Renman (1989) * Correspondence to: H. Tømmervik, Norwegian Institute for Nature Research (NINA), N-9296 Tromsø, Norway. E-mail: [email protected]

Copyright # 2010 John Wiley & Sons, Ltd.

claims that frequent use of vehicles in mountain areas is destructive for all vegetation types. However, the damage effects and the resistance vary from one vegetation type to another. It is regarded that vegetation types like fens (mires), exposed ridges, damp heaths, mountain meadows and vegetation types rich in lichens are especially vulnerable to terrain vehicles. Also forest vegetation types on deep, peaty subsoil have low resistance to damage (Renman, 1989; Kevan et al., 1995; Eastes et al., 2004).When evaluating the vulnerability to different vegetation formations, three main considerations are to be taken into account; extent of damage, resistance and ability to recover. Vegetation types on calcareous-rich, high-productive subsoil has high recover potentials, while for slow growing lichen species it will take years to recover. Also the extent of damage varies from a single track through a fen community to heavy damage were the vegetation cover is totally destroyed. In these cases secondary effects from water, wind and soil solifluction can deteriorate the initial damage dramatically (Renman, 1989). Kevan et al. (1995) showed that, in general, all tracks, regardless of age, showed small increases in the depth of thaw beneath them (c. 2.8 cm). The vegetation cover was

H. TØMMERVIK ET AL.

Figure 1. Map of the military exercise and battlefield areas in Mauken, Bla˚tind and Setermoen-Bardu, Troms County, northern Norway. This figure is available in colour online at wileyonlinelibrary.com

significantly reduced on all tracks, which had suffered multiple passages. In few sites where single passages were recorded, cover increased through proliferation of the sedge, Kobresia myosuroides. Experiences from other projects dealing with the detection of objects have shown that image resolution is the key factor for portraying objects at the smallest scale or detection of damage in the terrain (Ridley et al., 1997; Teterukovskiy, 2003; Eastes et al., 2004; Allard et al., 2004; Lechner et al., 2009). Accordingly, the high-resolution satellites as Ikonos-2 and Quickbird which can provide data with a spatial resolution up to 1 m and 0.6 m, respectively, are well suited for this purpose. For instance high-resolution satellite photographs offer a new picture of the tracks along which the Easter Island giant statues were hauled from the central quarry to the exhibition sites. The survey traced 32 km of seven major roads, confirmed by features on the ground (Lipo and Hunt, 2005). At this resolution scale most tracks from the aforementioned vehicles are visible. In addition also use of high-resolution aerial digital ortho-photos with 3 visual and 1 NIR–band is a very good tool in detection of ATVs-tracks (Allard et al., 2004). Forbes et al. (2009) used high-resolution imagery (Quickbird-2) in detection of off-road vehicle tracks due to oil and gas exploration in Yamal-Nenets in Siberia (Russia). They detected approximately 2500 km of vehicle tracks which occupied about 25 km2 as of 2005, and this show increasing human impacts as such elsewhere in Arctic. Although the problems ATVs are causing the environment are increasing; however, few studies (Eastes et al., 2004, Forbes et al. 2009) have been carried out using satellite remote sensing techniques in order to detect and map such damage. The objectives of the following study were to assess Copyright # 2010 John Wiley & Sons, Ltd.

the feasibility of using satellite imagery with high-spatial resolution to detect tracks from different ATVs and armoured vehicles and establish efficient methods to carry out a large-scale detection and mapping of all tracks and damaged areas within the study area. STUDY AREA The location of the study area and the battlefields is presented in Figure 1. In this project following three army battlefields were investigated with respect to damage caused by ATVs: Mauken (40.8 km2), Bla˚tind (140.9 km2) and Setermoen (152.3 km2). Within the Mauken exercise area (40.8 km2), mountain heaths and mountain meadows constitute more than half of the total area (Johansen (2009). Other landscape features within the area using the vegetation map of Johansen (2009) were birch forests (9.7 km2), fens (1.0 km2) and sparsely vegetated areas in the mountain region (5.5 km2). In the Bla˚tind and Setermoen areas the corresponding values were: Bla˚tind: birch forests 48.3 km2, fens 10.6 km2, mountain heaths/meadows 38.1 km2, areas with sparsely vegeation 43.9 km2; Setermoen: birch forests 55.8 km2, fens 7.0 km2, mountain heaths/meadows 35.4 km2 and sparsely vegetated areas 54.1 km2. DATA Satellite Data Ikonos-2 data were available for the Mauken area, while for the Setermoen and Bla˚tind study areas data from the Quickbird satellite was used. The Ikonos-2 image was acquired 2000-0815, the Quickbird images were acquired 2002-07-11 and 200307-19, respectively. See Table I for details. LAND DEGRADATION & DEVELOPMENT (2010)

DAMAGE CAUSED BY ATVS TO THE ENVIRONMENT

area. Also digital topographic maps as well as terrain models were utilized in order to correct the satellite images. METHODS

Figure 2. The all-terrain vehicles (ATVs) represented in the study from top to bottom: light ATV (bike), medium ATV: BV206 all-terrain carrier Alvis Ha¨gglunds and armoured vehicle CV90 Alvis Ha¨gglunds. This figure is available in colour online at wileyonlinelibrary.com

Other Data A digital vegetation map (Johansen, 2009) based on Landsat 7 ETMþ from the area was utilized in order to analyse the damage (in km2) on the main vegetation types within the Copyright # 2010 John Wiley & Sons, Ltd.

Geometric Correction of the Satellite Data The pan-sharpened Ikonos-2 image covering the study area Mauken was due to financial reasons ordered with imaging geometry. Due to lack of Rational Polynomial Coefficients (RPC) it was impossible to convert the image into orthophoto standard with respect to geometric accuracy. To obtain a best possible geometrical accuracy, the image was geometrically corrected based on control point collection and ‘warping’ techniques. The vector data having most reliable geometry within the study area is the Norwegian Mapping Authorities road database. Traditional maps, scale 1–50 000, delivered by the mapping authorities, do not correspond to the accuracy level inherent in the 1 m resolution Ikonos-2 image. Due to collection of control points from the road-database not covering mountain areas, a content geometric accuracy of the Mauken image was not obtained. The Pan-sharpening merging technique does not change at all the statistical parameters of the original images (Nikolakopoulus, 2005). Also for the Quickbird image delivered for covering the Bla˚tind area, problems with respect to geometric accuracy were experienced. The image was delivered with a geometriccorrection where a coarse-scaled DEM had been applied. An improvement using RPC’s was because of that not possible. Accordingly, also in this case a further geometric-correction of the image was performed using ground control points and warping techniques. The image covering Setermoen area was delivered separately as a panchromatic image (0.6 m) and a multi-spectral image with a ground resolution of 2.6 m. This image was delivered with all additional information (RPC) needed to perform geometric-correction with high accuracy. By combining the panchromatic and the multi-spectral image a pan-sharpened image was produced with a ground resolution of 0.6 m. Based on field verifications of the three mentioned images we can conclude that for both the Mauken and Bla˚tind areas an optimal geo-correction was not obtained. However, the geometric accuracy is by all means high enough to quantify the damage and relate to the 30  30 m2 spatial resolution Landsat based vegetation map. For the Quickbird image covering Setermoen area, geometric accuracy on a sub-pixel level was obtained. Detection Procedures Automatic detection is the best solution for object recognition within large datasets, or as in this case searching for vehicle tracks (Teterukovskiy, 2003). However, it is difficult to make automatic procedures with high-detection rates and low error rates, so visual interpretation techniques were used for detection of tracks in the satellite images. Prior to the extraction process LAND DEGRADATION & DEVELOPMENT (2010)

H. TØMMERVIK ET AL.

Table I. Satellite data specifications used in this study Area

Mauken Bla˚tind Setermoen Setermoen

Satellite

Bands

Spatial resolution (metres)

Geometry

Ikonos-2 Quickbird Quickbird Quickbird

3 (G,B og NIR) 4 (RGB þ NIR) 4 (RGB þ NIR) 1 Panchromatic

1 0.72 2.53 0.63

Not georeferenced Partly georeferenced Ortho ready Ortho ready

pan-sharpened 3-band images were composed based on the infrared channel and two of the visible channels. To obtain the ‘best possible’ images for track extraction purposes the 3-band colour composite images were treated with different types of image enhancements such as edge-detection, contrast stretch, histogram-equalization and different filtering techniques as a high boost filter (Gonzalez and Woods, 1992). Forbes et al. (2009) also used visual detection in order to digitize the off-road vehicle tracks in Siberia.

Acquistion date 15/8 11/7 19/7 19/7

2000 2002 2003 2003

‘Off nadir’ angle in degrees 30 14 9 9

into the vegetation map produced by Johansen (2009) in order to estimate the vegetation types influenced and damaged by the ATVs. The occurrence of overall vegetation types (forest, fen, heather, sparsely vegetated areas) within the whole test area was compared with corresponding occurrences within the influence area. The vegetation type nomenclature follows Fremstad (1997). Fieldwork

Development of Influence Zones and Impact Zones Initial inspection of the satellite images revealed a network of parallel and ‘overcrossing’ tracks. In order to infer the damage to the environment caused by all these tracks we decided to establish zones of damage and influence on both sides of the identified tracks: 10 m damage and influence zone in where there are extensive damage and/influence. A zone of 25 m was also established and included damaged area (tracks and battle field areas) as well as indirect influence (e.g. alterations of the water level in the fens). Finally, a zone of 50 m, which in addition to damage and influence areas, also include impact/influence zones concerning wildlife (e.g. birds, reindeer, small mammals and mice) was established (Tømmervik et al., 2005). All observed tracks were digitized by hand using the ENVI software, and the influence zones on each side of the tracks were applied. The influence zones were then superimposed

Additional to remote sensing detection of tracks, field inspections were performed to verify the satellite image observations and to evaluate ecological effects of the vehicle tracks crossing the different vegetation types. Evaluations regarding preferences in use of different vegetation types for driving were also performed, as well as registration of category and degree of observed damage within the different vegetation types.

RESULTS Enhancements of Satellite Images In Figure 3 we present an original pan-sharpened Quickbird image which is filtered with a high boost routine. We clearly see the enhanced detection and identification possibilities. The investigation revealed a network of all-terrain tracks in fens, mountain heaths, ridges and forests, but the damage in

Figure 3. Left: original pan-sharpened Quickbird image from Setermoen shooting range. Middle: high boost filtering and 2 per cent linear stretching. Right: pan-sharpened Quickbird image from Setermoen shooting range. High boost filtered. This figure is available in colour online at wileyonlinelibrary.com Copyright # 2010 John Wiley & Sons, Ltd.

LAND DEGRADATION & DEVELOPMENT (2010)

DAMAGE CAUSED BY ATVS TO THE ENVIRONMENT

Figure 4. Single track from ATVs (6-wheeled motorbike) in a fen can be observed in the Ikonos satellite image. Also observed is a road under construction. This figure is available in colour online at wileyonlinelibrary.com

the more dry vegetation types were not as distinct as in the fens. Detection of Tracks The pictures shown in Figure 4 illustrate to what extent different types of ATVs vehicle tracks are detectable in highresolution satellite images. Even a single track (Figure 4 upper left) caused by a small ATVs which had crossed over a mire dominated by grass and sedges could be detected in the Ikonos-2 image. This type of track is not easy to detect using

traditional aerial photographs (black/white). However, in multi-spectral satellite images the contrast between the densely vegetated grass/sedges and the track line is clearly portrayed. In Figure 5, several vehicle tracks from medium–heavy ATVs on a mountain ridge are visualized, and some of these tracks were older than 25 years. These types of tracks are also easily detected in the high-resolution satellite images due to the contrast between damaged vegetation along the track line and the surrounding healthy vegetation. Even

Figure 5. Ridge with several parallel tracks from medium heavy ATVs in Bla˚tind. Left: in situ photo and right: Quickbird image over the same area. Tracks are easily seen as dark stripes in the satellite image. This figure is available in colour online at wileyonlinelibrary.com Copyright # 2010 John Wiley & Sons, Ltd.

LAND DEGRADATION & DEVELOPMENT (2010)

H. TØMMERVIK ET AL.

Figure 6. Quickbird image of tracks caused by heavy armoured vehicles in Setermoen battlefield. The rectangle indicates the observation place for the in situ picture to the right. This figure is available in colour online at wileyonlinelibrary.com

though these vehicles are constructed to float on top of a snow-covered surface, the ground pressure is heavy enough to damage the vegetation. In Figure 6, tracks caused by heavy armoured vehicles (Leopard II and CV-90 and other heavy vehicles) in bog at Varden area, Setermoen, are illustrated. The photo on left shows the heavy damage mainly done by armoured vehicles like Leopard and CV-90, and in the Quickbird image at the left we clearly see the tracks as the dark areas in the upper half of the image. Even the two bands in the track are visible. The tracks are so marked that they are easily visible in the satellite images, and could probably also have been identified in medium resolution satellite images. In Table II we present the characteristics of the tracks from the different ATVs and we can here observe the variation in width and depth of the different type of tracks. The accuracy in detection of tracks was assessed and this assessment revealed that 62 of 424 tracks (Table III) investigated in field were not detectable on the imagery, meaning an accuracy of

85 per cent. The most of these tracks not detectable on the images were located in the forests (Table III), while some of them were located as single tracks of light ATVs on mountain heaths and in forests. The vast majority of tracks located in fens (90 per cent) were detected in the images. If we only take ‘new’ tracks and also exclude single tracks in the accuracy analysis the overall accuracy is estimated to 91 per cent. Length of Tracks and the Areas of Influence Zones and Impact Zones The total length of the tracks within the three battlefields was calculated to be ca. 1000 km (Table IV). Figure 7 shows the digitized vehicle tracks in the battlefield areas superimposed on the satellite images. The areas inside the influence-zone (50 m from each side of the track-line) are for the battlefield in Bla˚tind summarized to 23 km2 of a total area of 140.9 km2 (Table IV). This means that 17 per cent of the whole battlefield area in Bla˚tind is influenced by vehicle tracks. Corresponding

Table II. Characteristic traits of tracks from the different ATVs active in from the study areas Type of ATVs

Average width (m) Variation in width (m) Average in depth (cm) Variation in depth Cover of vegetation (per cent) Cover of bare soil (per cent) Cover of bare rock (per cent) Number of sites

Leopard I/II (armoured)

CV-90 armoured vehicle

M113 armoured vehicle (old tracks)

Medium large ATV: BV-202 (old tracks)

Medium large ATV: BV-206

Small ATV (bikes)

5 4.0–10.0 40 30–70 35 50 15 7

4 3.5–6.0 40 30–50 45 50 5 23

2.7 2.5–2.8 14 10–20 97 3 – 27

2.5 1.9–15,0 16 0–40 76 8 16 268

2.4 2.1–2.7 12 0–30 73 25 2 42

1.7 1.5–2.0 25 25 15 30 55 57

Copyright # 2010 John Wiley & Sons, Ltd.

LAND DEGRADATION & DEVELOPMENT (2010)

DAMAGE CAUSED BY ATVS TO THE ENVIRONMENT

Table III. Accuracy table due to detection of the tracks within the main land cover types: mountain heaths, fens and forests. The majority of M113 tracks are more than 25 years old CV-90 M113 Armoured Medium Medium Small ATV Small ATV Overall Leopard vehicle large ATV: BV-202 large (bikes) (bikes); accuracy Armoured Armoured vehicle (old tracks) (old tracks) ATV: BV-206 Single tracks Vehicle (tank) per cent per cent per cent per cent per cent per cent per cent per cent

Type of ATVs

Mountain heaths Fens (mires) Forests Overall accuracy Number of sites

100 100 100 100 7

100 100 100 100 23

43 78 50 60 27

86 91 84 88 268

values for Mauken battlefield area are a total area of 40.8 km2 with an influence area of 8.9 km2, or 21.8 per cent. The total area of the Setermoen is 152.3 km2, and here was the influence zone of 50 m estimated to 23.9 km2, or 15.7 per cent. Infrastructure elements like artillery- and gun stands, buildings, etc. in the three battlefields were estimated to 2.5 km2, while the area of roads was estimated to 5.9 km2 (Table IV). Vegetation Influenced or Damaged Our first impression in the initial stage of the project was that fen and mire vegetation is commonly used to drive on by ATVs. To verify this impression vegetation maps were applied for all the test areas and the extent of the influence zones superimposed into the vegetation maps (Figure 8) and the areas calculated (Table V). The overall damage and influence within the fen (mire) vegetation was estimated to 13 per cent within the 10 m zone of influence, and 38 per cent within the 50 m zone of damage/influence, respectively (Table V), but the damage and influence varied a lot amongst the fen and mire types, and here the K2a poor hummock fen was most damaged/influenced (Table V). Not only the vegetation types with highest biodiversity, low and tall herb forests, experienced considerable loss and influence (Table V), but also vegetation types with thin soils above the bedrock (e.g. lichen woodland and different vegetation types on ridges experienced significant losses in the different

80 92 86 89 42

80 90 71 86 37

20 80 20 50 20

79 90 76 85 424

zones (Table V), and the tracks and wear in these areas were also clearly visible (Figure 5).

DISCUSSION Development of Influence Zones and Impact Zones Our approach using high-resolution satellite imagery has shown that we can detect all terrain vehicle tracks with a sufficient accuracy. The accuracy of the Ikonos-2 image was lower than the Quickbird imagery due to the acquisition angle and problems with the geometry as well as lowered spatial resolution. The probability of detecting a pixel belonging to a track depends on how the pixel grey levels differ from the grey levels of pixels in the neighbourhood and on additional prior information. Teterukovskiy (2003) used a Bayesian approach in order to detect tracks in aerial photos and he based the inference on an a priori knowledge of the structure of tracks. Teterukovskiy (2003) used the ‘Gibbs sampler’ procedure in order to construct the most probable pattern or occurence of tracks in the area. The method is applied to aerial photos with cell size of 1  1 m2, but Teterukovskiy (2003) applied the same methodology to SPOT panchromatic images with cell size of 10  10 m2 for detection of larger tracks. The results were similar taking into consideration the coarser resolution. Even for detection of trails of width comparable width or smaller than the cell size, positive results can be achieved. Therefore we argue

Table IV. Length of tracks and length/area of roads and infrastructure within the battlefields Battlefield

Area Length of Length of Area within Area within Area within Area of roads Area of Total area the 10 m roads in km2 tracks the 25 m the 50 m in km2 infra- structure of infra-structure influence in km in km influence influence n km2 in km2 zone km2 zone km2 zone km2

Mauken 40.78 Bla˚tind 140.94 Setermoen 152.32 Total 334.04

140 364 500 1004

77.9 143.3 73.4 294.6

Copyright # 2010 John Wiley & Sons, Ltd.

2.6 6.9 7.8 17.7

5.5 14.5 15.3 35.3

8.9 23.3 23.9 55.7

1.56 2.87 1.46 5.89

0.40 1.30 0.75 2.45

1.96 4.16 2.22 8.34

LAND DEGRADATION & DEVELOPMENT (2010)

H. TØMMERVIK ET AL.

Figure 7. Digitised tracks in Bla˚tind battlefield. Blue line indicates the border of the battlefield area. This figure is available in colour online at wileyonlinelibrary.com

that this methodology might be a candidate for more efficient detection of tracks in any remotely sensed image and probably better than the results we present. The question of assessing the quality, however, is not discussed by Teterukovskiy (2003). In many practical applications that deal with detecting objects, the exact shape of the object is of less importance than the location. Depending on the problem, it can be either more important not to erroneously detect the objects that do not exist than to miss the objects that do exist or vice versa. This type of trade off is well-

Figure 8. The influence zones of 50 m superimposed into the vegetation map over the Bla˚tind battlefield. This figure is available in colour online at wileyonlinelibrary.com Copyright # 2010 John Wiley & Sons, Ltd.

known in statistics as errors of the first and of the second kind. An alternative method is to increase the spatial resolution of imagery that is too low to meet requirements of detection of small objects and linear features with low widths. A super-resolution mapping method (Thornton et al., 2006) was found to work well for objects approximately as wide as a pixel; however, objects lesser than a pixel width had low classification accuracy. Whilst Tatem et al.’s (2002) super-resolution technique recreated sub-pixel spatial pattern class proportions but did not map the true location of sub-pixel features. Other methods which focus on extracting linear features are predominately being used to extract roads to update GIS databases from imagery (Quackenbush, 2004) rather than linear tracks set by ATVs in forests, mires and mountain heaths. Additionally, linear feature extraction techniques ignore other nonlinear features thus having limited application. Lechener et al. (2009) analysed and quantified the effects of patch size, length, grid position and detectability of small and linear vegetation features in land cover mapping, and the methods here are of interest. They found that mapping error was highest when the scale of the linear or the square feature and the raster grid coincided, and they showed that the spatial resolution of the grid should be many times finer in order to extract these features accurately. For example, a square shaped patch needed an area of at least 11 pixels to achieve a mean accuracy of 75 per cent, whilst a linear patch with a width to length ratio of 4 needed an area of 12 pixels. In our case, the contrast from the damaged vegetation or denuded soils/ substrates, help us in detection of linear features since pansharpened Quickbird and Ikonos imagery (spatial resolution: 0.63/0.72 and 1 m), enabled us to detect single tracks with a width less than 2 m. Also old tracks (> 25 years) were detectable, especially on thin and wet vegetation, and this is due to wear and erosion in the trails (Tømmervik et al., 2005). Poulis and You (2010) presented a novel vision-based system for automatic detection and extraction of complex road networks from various sensor resources such as aerial photographs, satellite images and LiDAR. Uniquely, the proposed system is an integrated solution that merges the power of perceptual grouping theory (Gabor filtering, tensor voting) and optimized segmentation techniques (global optimization using graph-cuts) into a unified framework to address the challenging problems of geospatial feature detection and classification. The use of specialised segmentation techniques considerably improved the segmented road pixels and extensive tests showed that the proposed system performs well for all data types and scenes, and has consistently achieved a minimum success rate of an average of 69.3 per cent. However, the segmentation process is based on the assumption that the Gaussian distributions of the foreground and background pixels (i.e. road and nonLAND DEGRADATION & DEVELOPMENT (2010)

DAMAGE CAUSED BY ATVS TO THE ENVIRONMENT

Table V. Total area of the vegetation types within the battlefields and the area of each vegetation types within the zones of damage and influence of 10, 25 and 50 m of each side of the tracks, respectively. The source for the extraction of zones of damage and influence is the vegetation map made by Johansen (2009). The vegetation type nomenclature follows Fremstad, 1997 Different zones of influence along the tracks Total area

Zone 10 m

Zone 25 m

Zone 50 m

No.

Vegetation types

km2

per cent

km2

per cent

km2

per cent

km2

per cent

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28

A2c pine forest A2c open pine forests with lichen A2c mixed forest A1b lichen woodland A4c crowberry woodland A4a bilberry woodland A5c small-fern woodland C2a tall-herb downy birch forest C2c low-herb downy birch forest K2a poor hummock fen L2 /L4 intermediate lawn fen L1 intermediate fen with willows Carpet/mud bottom fen Bottom fen/swamp R1a alpine ridge; trailing azalea R1a/R3c mountain avens ridge R2b dwarf birch lichen ridge S3 bilberry/crowberry heath S2a juniper/dwarf birch heath R5 graminoid ridge S6-S7 alpine tall-herb meadows T1-T3 snow patch vegetation T4 dwarf willow snow patch T5 poor bryophyte meadow R6 wood-rush ridge Snow and glacier Lakes, rivers and ponds Barrens, rocks Total

5.7 0.6 3.5 0.2 4.3 4.8 36.2 42.9 14.2 1.2 9.8 6.1 0.8 0.1 0.0 8.2 0.0 5.7 31.3 13.0 25.3 10.1 24.7 13.9 14.5 44.5 5.7 6.3 334.0

1.7 0.2 1.0 0.1 1.3 1.4 10.9 12.9 4.3 0.4 2.9 1.8 0.2 0.0 0.0 2.5 0.0 1.7 9.4 3.9 7.6 3.0 7.4 4.2 4.3 13.3 1.7 1.9 100.0

0.3 0.2 0.3 0.1 0.9 0.4 2.3 3.1 0.5 0.3 1.2 0.8 0.1 0.0 0.0 0.6 0.0 0.5 3.0 0.7 1.1 0.2 0.4 0.4 0.1 0.2 0.0 0.0 17.7

5.3 33.3 8.6 50.0 20.9 8.3 6.4 7.2 3.5 25.0 12.2 13.1 12.5 0.0 0.0 7.3 0.0 8.8 9.6 5.4 4.3 2.0 1.6 2.9 0.7 0.4 0.0 0.0 5.3

0.5 0.3 0.6 0.1 1.6 0.6 4.7 6.3 1.0 0.6 2.4 1.5 0.2 0.0 0.0 1.2 0.0 1.1 6.4 1.5 2.2 0.6 0.7 0.6 0.2 0.4 0.0 0.0 35.3

8.8 50.0 17.1 50.0 37.2 12.5 13.0 14.7 7.0 50.0 24.5 24.6 25.0 0.0 0.0 14.6 0.0 19.3 20.4 11.5 8.7 5.9 2.8 4.3 1.4 0.9 0.0 0.0 10.6

1.0 0.4 0.9 0.1 2.2 0.9 8.0 10.6 1.8 0.7 3.4 2.4 0.3 0.0 0.0 1.7 0.0 1.7 9.8 2.2 3.5 0.9 1.1 1.0 0.3 0.8 0.1 0.1 55.9

17.5 66.7 25.7 50.0 51.2 18.8 22.1 24.7 12.7 58.3 34.7 39.3 37.5 0.0 0.0 20.7 0.0 29.8 31.3 16.9 13.8 8.9 4.5 7.2 2.1 1.8 1.8 1.6 16.8

road pixels) are easily separable. Although this is true in most cases with colour imagery, there are instances where non-road pixels exhibit the same reflectance properties (i.e. have the same color) as the road pixels. In our case dealing with often subtle tracks from light ATVs, such a visionbased system may have helped, but we also encountered problems with networks of small brooks (creeks) that could be misinterpreted as ATVs tracks, and an automatic procedure may have detected this network of brooks as ATVs tracks. However, tracks from armoured ATVs, could be more easily detected by such a vision based system for automatic detection, but taken in consideration that we obtained overall classification accuracy of 90 per cent of ATVs tracks (including light ATVs) in the most vulnerable type of vegetation (fens), we consider the visual and manual method as the best for present time. The real-life problem that motivated this study was to monitor the potential damage to the terrain induced not only by use of ATVs in military battlefields, but also civilian use Copyright # 2010 John Wiley & Sons, Ltd.

of ATVs has increased. The use of civilian vehicles is only allowed in established (legal) tracks or corridors followed by illegal use in other areas (Tømmervik et al., 2005; Strann, 2010). These areas are very large and therefore hard to inspect and monitor. This means that the cost of erroneous detection of the non-existing track is high. On the other hand, the exact shape of the tracks is not of serious importance, and it is acceptable for the results of the detection to have gaps. Therefore our prior concern was to detect the objects that most definitely were the tracks. If the opposite would be the case, e.g. if it would be important to detect all tracks with the risk of erroneously detecting something else like ridge edges or creeks, then the threshold value in the ‘edge detection’ procedures must be lowered. Eastes et al. (2004) tested other sensors and wavelengths in order to detect tracks from military vehicles. Their results suggest that remotely sensed thermal signatures could differentiate tracked and wheeled vehicles on terrain in many areas of the world of strategic interest. LAND DEGRADATION & DEVELOPMENT (2010)

H. TØMMERVIK ET AL.

Vegetation Influenced or Damaged These investigations showed that the mire vegetation was the category that was most damaged or influenced by a large network of all-terrain vehicle tracks (Table V). Two main conclusions can be outlined from these comparisons: (1) Fen vegetation are over-represented in the influence areas for all the test sites, while (2) vegetation types in the mountain region are underrepresented inside the influence areas in all the test sites. The largest damage on mire vegetation was registered in Mauken and Setermoen, while the fens in Bla˚tind were less damaged. In mire vegetation a decrease of especially moss species was observed within the all-terrain tracks. The traffic with the all-terrain vehicles had led many places to a transformation from typical mire and fen vegetation (drier conditions) to a carpet/mud bottom vegetation influenced by high water levels (Tømmervik et al., 2005), and hence this can cause trouble for waders and other birds living in these habitats (Strann, 2010). The regional investigation also revealed a network of all-terrain tracks in mountain heaths and ridges, but the damage was not as distinct as in the fens. The effect on the biodiversity was large in areas with thin soils above the bedrock (Tømmervik et al., 2005), and the tracks and wear in these areas were clearly visible and detectable by use of the high-resolution imagery. CONCLUSIONS We have by the use of high-resolution satellite images (Quickbird and Ikonos) investigated three exercise and shooting ranges (battlefields) for the Norwegian Army with respect to tracks and damage caused by ATVs. Applying simple image processing techniques our investigation showed that even small tracks with minor damage on the vegetation caused by light ATVs were easily detected in the satellite images. Also old tracks (> 25 years) caused by medium to heavy ATVs were detected. Application of more automatic methods, however, might ease the detection and monitoring in the future. The length of the all-terrain vehicle tracks within the three main military battlefields was estimated to more than 1000 km, and this has influenced vulnerable vegetation types within the study areas considerably. We conclude that high-resolution optical satellite imagery is well suited for surveying damage on the vegetation caused by terrain vehicles in northern Norway, even damage caused by light and single ATVs.

ACKNOWLEDGEMENTS The present study was financed by the Norwegian Defence Estates Agency, the Strategic Institute Program REMA and by internal funding from Norwegian Institute for Copyright # 2010 John Wiley & Sons, Ltd.

Nature Research (NINA) and Northern Research Institute (NORUT).

REFERENCES Allard A, Lo¨fgren P, Sundquist S. 2004. Skador pa˚ mark och vegetation i de svenska fja¨llen till fo¨ljd av barmarksko¨rning. Report 126, Swedish University for Agricultural Sciences: Umea˚, Sweden. Eastes JW, Mason GL, Kusinger AE. 2004. Thermal signature characteristics of vehicle/terrain interaction disturbances: Implications for battlefield vehicle classification. Applied Spectrosccopy 58: 510–515. Forbes BC, Stammler F, Kumpula T, Meschtyb N, Pajunen A, Kaarlejærvi E. 2009. High resilience in the Yamal–Nenets social–ecological system, West Siberian Arctic, Russia. Proceedings of National Academy of Sciences (PNAS) 106: 22041–22048. Fremstad E. 1997. Vegetasjonstyper i Norge. (Vegetation types in Norway). NINA Temahefte 1: Trondheim, Norway. Gonzalez RC, Woods RE. 1992. Digital Image Processing, 2nd edn. Addison-Wesley Publishing Company: Boston, MA. Johansen B. 2009. Vegetasjonskart for Norge basert pa˚ Landsat TM/ ETM R data. Report 4/2009, Northern Research Institute: Tromsø, Norway. Kevan PG, Forbes BC, Kevan SM, Behanpelletier V. 1995. Vehicle tracks on high arctic tundra–their effects on the soil, vegetation, and soil anthropods. Journal of Applied Ecology 32: 655–667. Lechner AM, A. Stein A, Jones SD, Ferwerda JG, 2009. Remote sensing of small and linear features: Quantifying the effects of patch size and length, grid position and detectability on land cover mapping. Remote Sensing of Environment 113: 2194–2204. Lipo CP, Hunt TL. 2005. Mapping prehistoric statue roads on Easter Island. Antiquity 79: 158–168. Nikolakopoulus KG. 2005. Eight different fusion techniques for use with very high-resolution data. SPIE Europe International Sympsoium Remote Sensing, Technical Programme and Abstracts, Bruges, Belgium, 19–22 September 2005; 86. Norberg MBE, Simons S, Alm T, Alsos IG, Jakobsen NP, Lund L. 1998. Prosjekt ‘Revegetering Karasjokfjellet’ 1995–1997. Tromura naturvitenska 83: 1–44. Poullis C, You S. 2010. Delineation and geometric modeling of road networks. ISPRS Journal of Photogrammetry and Remote Sensing 65: 165–1181. Quackenbush LJ. 2004. A review of techniques for extracting linear features from imagery. Photogrammetry Engineering and Remote Sensing 70: 13831392. Renman G. 1989. Barmarksko¨rning i fja¨llen. Effekter av ko¨rning med terra¨nghjulingar pa˚ mark och vegetation. Rapporter fra˚n SNV, 071SNV. Statens Naturva˚rdsverk: Solna, Sweden. Ridley HM, Atkinson PM, Aplin P, Muller JP, Dowman I. 1997. Evaluating the potential of the forthcoming commercial U.S. high resolution satellite sensor imagery at the ordnance survey. Photogrammetry, Engineering and Remote Sensing 63: 997–1005. Strann KB. 2010. En smak av Sibir. Blant mygg og østlige vadefugler i Indre Finnmark. Ottar 1: 11–17. Tatem AJ, Lewis HG, Atkinson PM, Nixon MS. 2002. Super-resolution land cover pattern prediction using a Hopfield neural network. Remote Sensing of Environment 79: 114. Thornton MW, Atkinson PM, Holland DA. 2006. Sub-pixel mapping of rural land cover objects from fine spatial resolution satellite sensor imagery using super-resolution pixel-swapping. International Journal of Remote Sensing 27: 473491. Teterukovskiy A. 2003. Detection of tracks in aerial photos by the Gibbs sampler. International Journal of Pattern Recognition 17: 1–16. Tømmervik H, Erikstad L, Jacobsen KO, Strann KB, Bakkestuen V, Aarrestad PA, Yoccoz NG, Hagen D, Johnsen TV, Johansen B, Høgda KA, Ahmed SH, Dahl R, Bargel TH, Olsen L. 2005. Langtidsvirkninger pa˚ naturmiljøet av Forsvarets virksomhet i Troms. NINA Rapport, Vol. 49, Tromsø, Norway. LAND DEGRADATION & DEVELOPMENT (2010)