1. High-resolution vegetation maps are a valuable resource for conservation, land management and research. In Great Britain, the National Vegetation Classification (NVC) is widely used to describe vegetation communities. NVC maps are typically produced from ground surveys which are prohibitively expensive for large areas. An approach to produce NVC maps more cost-effectively for large areas would be valuable.
2. Creation of vegetation community maps from aerial or satellite images has often had limited success as the clusters separable by spectral reflectance frequently do not correspond well to vegetation community classes. Such maps have also been produced by exploring correlations between community occurrence and environmental variables. The latter approach can have limitations where anthropogenic activities have altered the distribution of vegetation communities. We combined these two approaches and classified 24 common NVC classes of the Yorkshire Dales and an additional class ‘wood’ consisting of trees and bushes at a resolution of 5 m from mostly remotely sensed variables with the algorithm random forest.
3. Classification accuracy was highest when environmental variables at low and high resolution (50 and 5–10 m, respectively) were added to aerial image information aggregated to a resolution of 5 m. Low-resolution environmental variables are likely to be correlated with the dominant vegetation surrounding a location and thus could represent critical area requirements or local species pool effects, while high-resolution environmental variables represent the environmental conditions at the location.
4. Overall classification accuracy was 87–92%. The median producer’s and user’s class accuracies were 95% (58–100%) and 92% (67–100%), respectively.
5.Synthesis and applications. The classification accuracies achieved in this study, the number of classes differentiated, their level of detail and the resolution were high compared with those of other studies. This approach could allow the production of good-quality NVC maps for large areas. In contrast to existing maps of broad land cover types, such maps would provide more detailed vegetation community data for applications like the monitoring of vegetation in a changing climate, the study of animal–habitat relationships, conservation management or land use planning.
The accuracy of a vegetation map may also depend on the resolution of the variables used in its prediction. In image interpretation, Cushnie (1987) and Wardley, Milton & Hill (1987) found that prediction accuracies of some classes improved as the image resolution was reduced owing to noise reduction in the spectral signals. In species distribution models, the relationship between alpine vegetation and topographical variables was influenced by the resolution of the topographical variables (Gottfried, Pauli & Grabherr 1998). The occurrence of species may depend not only directly on environmental variables but also on factors intrinsic to the species or community, such as dispersal and competition (Leathwick 1998; Miller, Franklin & Aspinall 2007). Low-resolution environmental variables may be correlated with the dominant vegetation. They may therefore represent the community most likely to succeed through intrinsic factors, while high-resolution environmental variables may better represent the actual conditions at the location.
In Great Britain, the National Vegetation Classification (NVC) is a standard for describing the natural, semi-natural and some artificial habitats (Rodwell 2000). It distinguishes vegetation to the level of communities, a much finer distinction than land cover types. The difference between communities can be small, with the same species being dominant in several distinct NVC communities (see Appendix S1 in Supporting Information). NVC community maps are typically created by ground surveys. Therefore, existing maps are usually of limited extent, and an approach that allows the creation of NVC community maps for large areas more cost-effectively would be valuable. Attempts to derive NVC community maps from environmental data and aerial and satellite images have previously often had limited success (Sanderson et al. 1995; Reid & Quarmby 2000; Shanmugam et al. 2003). We use a novel approach and integrate image interpretation with species distribution modelling using environmental variables at the local and neighbourhood scales to classify NVC communities with the algorithm random forest. In this paper, we classify NVC communities from a combination of readily available, colour aerial photography and environmental variables at a high resolution (5 m) for an area of the Yorkshire Dales National Park, an upland area in Great Britain. We calculate NVC class and patch characteristics from the resulting map. We also investigate the influence of using predictor variables at varying resolutions on prediction accuracy of NVC classes.
Materials and methods
The study was carried out in the Yorkshire Dales National Park (Fig. 1a), in the central Pennines of northern England. The National Park consists of valleys dominated by pastures and meadows, and plateau tops dominated by grass moorland, upland heathland and blanket peat mire.
Aerial images were obtained from existing stock of Bluesky International Limited (Copyright GeoPerspectives) by the Yorkshire Dales National Park Authority (YDNPA). The images consisted of digital scans of aerial photographs captured on film, had three bands (red, green and blue), a resolution of 0·25 m and were orthorectified and compatible with the Ordnance Survey MasterMap (http://www.ordnancesurvey.co.uk/oswebsite/products/os-mastermap/topography-layer/index.html). The images were photographed between 15 April and 18 December 2002. For each location, photography from only a single date was available. As the spectral characteristics of vegetation change during the year (Lillesand, Kiefer & Chipman 2008), we performed a separate classification for each season and grouped the images as follows: spring (15–17 April), which covered areas in the south of the study area, autumn (30 September and 1 October), which covered areas in the middle and far north of the study area, and winter (18 December), which covered areas in the north of the study area (Fig. 1b).
2 The ratios between the reflectance values of each band (hereafter referred to as ratio): They preserve the spectral characteristics of features, but are more independent of variations in illumination, for example whether a feature was located in sun or in shade at the time it was photographed (Price 1998; Lillesand, Kiefer & Chipman 2008).
3 A measure of heterogeneity of spectral reflectance (hereafter referred to as SD): classes can differ in their degree of heterogeneity (Cushnie 1987), which we utilized to help separate classes.
We compared the prediction accuracies between image data at 0·25- and 5-m resolution for the spring season and used the resolution with the best results for the other seasons. For the 5-m resolution, we followed the approach of Chapman et al. (2010). We calculated the means for the 5-m resolution for each band and calculated the ratios between the means. For the SD measure, we averaged the standard deviation of pixel values within 0·5 m for the 5-m resolution of each band. Hence, we calculated nine variables (three types of variable × three bands) from the aerial images. For the 0·25-m resolution, we used the reflectance values of each pixel as means as well as the ratios between the three bands. Zero reflectance values were replaced by 0·001 to allow the calculation of ratios. We could not calculate SD for the 0·25-m resolution analogue to the 5-m resolution as 0·25 m was the resolution of the aerial images. For seasons with more than one flying date (spring, autumn), we used flying date as an additional predictor variable.
We used elevation, aspect, slope and soil type as environmental predictor variables. They are readily available and can influence plant species distribution through, for example, rainfall, temperature, solar radiation and soil nutrient content (Austin 2002; Bennie et al. 2006). Soil type was extracted from the simplified version of the National Soil Map data set (NATMAP soilscapes) after conversion into a 5-m raster (NSRI 2009). The NATMAP soilscapes 1:125 000 vector map classifies soil types within England and Wales into 27 broad classes. Elevation was obtained as a mainly remotely sensed digital terrain model (DTM) at two resolutions: at 50-m resolution with a height resolution of 1 and 5-m height accuracy and at 10-m resolution with 1-m height resolution and 2·5- to 5-m height accuracy (Land-Form PANORAMA and PROFILE data downloaded from the EDINA Digimap OS service; http://edina.ac.uk/digimap). Aspect and slope were calculated at 50-m resolution from the 50-m DTM and aspect at 10-m resolution from the 10-m DTM with the ArcGIS – Spatial Analyst (ESRI Redlands, California, U.S.). Aspect was converted into a factor variable with nine levels (flat, north, north-west, west, south-west, south, south-east, east and north-east). Slope data at 5-m resolution derived from the NEXTMAP DTM (Intermap Technologies) with a root mean square error of 1 m was obtained from Natural England (http://www.naturalengland.org.uk) under the YDNPA’s Mapping Services Agreement.
Vegetation Community Data
We used sample NVC data from a large-scale survey of semi-natural habitats in the Yorkshire Dales in 2002–2006. The survey was not specifically designed for this study, but aimed to provide the YDNPA with a base map for their activities (see Appendix S2, Supporting Information for survey details). We used 24 NVC classes (see Appendix S1, Supporting Information), which, according to surveys and our knowledge of the study area (see Appendix S3, Supporting Information), were the most common. We removed the outer 10 m from each polygon to minimize bias from mapping errors. The median size of polygons containing only the above 24 NVC classes was 2411 m2. As usual for such a large-scale survey (see Appendix S2, Supporting Information), several NVC communities were sometimes identified in one area and then mapped as one mosaic polygon owing to time constraints and the mapping scale. In total, 1090 polygons covering 55 km2 consisted of mosaics of several of the above 24 classes. A further 385 polygons covering 5·0 km2 consisted of one or more of the above 24 classes in mosaics with scree, rock, bare earth, limestone pavement, shrubs or trees. Overall, 4381 polygons covering 114·2 km2 consisted of only one of the above 24 classes. As the mosaic polygons did not allow us to accurately ascertain class membership, we did not use them as training data. A further 362 polygons covering 6·2 km2 included classes additional to the above 24 classes and were not used. From the aerial images, we digitized areas of woodland and shrub covering 23·6 km2 to use as an additional class ‘wood’.
We converted the polygons into raster maps with a resolution of 5 and 0·25 m and randomly selected 1000 pixels per class per season to use as training pixels. This allowed us to run the classifications on a 3-GB-RAM computer while still using pixels from a geographically wide extent. One (in spring) and three (in winter) of the rarer NVC classes had fewer than 1000 training pixels available. The minimum number of training pixels per NVC class used was 255. For the 5-m resolution, training pixels per class and season came on average from 27 (median) polygons (range: 2–139) and spanned a minimum convex polygon of 170·7 km2 (median; range: 0·4–894·9 km2). Three NVC classes (CG2, M10 and M15) were considered rare in the north of the study area, the area photographed in winter, and were not included in this classification. Hence, we used 22 classes in winter and 25 in both autumn and spring. For the autumn classification, 25 000 pixels from 1219 NVC polygons and 399 wood polygons were used, for the spring classification, 24 894 pixels from 950 NVC polygons and 244 wood polygons, and for the winter classification, 20 515 pixels from 469 NVC polygons and 412 wood polygons.
We compared the classification results using different resolutions of image variables (0·25 and 5 m) and of environmental variables (5–10, 50 m and both combined). We also compared classification results using environmental or aerial image variables only or both combined.
For each of the many classification trees in random forest, a training data set is constructed by leaving approximately one-third of the original data ‘out-of-bag’ and sampling with replacement from the other two-third of the original data. Each tree is used to predict the ‘out-of-bag’ data, and a classification is assigned by majority vote. The error rate is the proportion of times that the predicted class is not the same as the true class. The ‘out-of-bag’ estimates are unbiased, and no external cross-validation is necessary (Breiman 2001; Prasad, Iverson & Liaw 2006; Cutler et al. 2007). Because of the random elements in the algorithm (see Appendix S4, Supporting Information), results vary slightly between repetitions of the same model. For example, 10 repetitions for our spring season produced a mean error rate of 9·06 ± 0·04% (SD). As the variation was low and computational costs of repetitions high, we report the results of one run only. We calculated overall accuracies as 100%–‘out-of-bag’ error rate. User’s accuracies estimate the probability that a predicted pixel of class A is truly class A and were calculated as follows: the number of correctly predicted pixels of class A/number of all pixels predicted as class A. Producer’s accuracies estimate the probability that a pixel of class A is predicted as such and were calculated as follows: the number of 7correctly predicted pixels of class A/number of training pixels of class A (Lillesand, Kiefer & Chipman 2008). We tested for residual spatial autocorrelation, which may affect the accuracy of predictions (Beale et al. 2010). We calculated the correlation between a distance matrix of sample locations (3000 randomly selected training pixels per season) and a distance matrix of the residuals (one if predicted correctly and zero otherwise) and used a Mantel test with 999 permutations.
We predicted an NVC map for the whole study area (1630 km2). Features recorded as water, buildings, manmade surfaces, gardens, roads, paths, tracks and railways in a topographical map (Ordnance Survey MasterMap) were masked out. ‘Obstructing features’ in the MasterMap (features that were mainly field walls) were buffered by 1 m and masked out. The percentage of each class in the predicted map was calculated. Further summary statistics were calculated in Fragstats 3.3 (McGarigal et al. 2002) after resampling the predicted map to a resolution of 10 m by majority rule. In this map, features recorded as roads, paths, tracks, railways and walls were not masked out, so that patches were not artificially divided into smaller patches. We calculated the number of patches, mean patch area and a clumpiness index denoting spatial aggregation of classes. The clumpiness index is zero for a random distribution of patches, −1 for maximally disaggregated patches and +1 for maximally clumped patches (McGarigal et al. 2002).
Classifications using image data (means and ratios, calculated as the mean of reflectance values per band and their ratios) photographed in spring had higher accuracies for the 5-m resolution than for the 0·25-m resolution (Table 1). Using only image data (5-m resolution), overall classification accuracies were highest in all seasons when all three types of image variables [means, ratios and SDs (heterogeneity of reflectance per band)] were used. However, when only using these image data, overall accuracies were poor (39–52%). Accuracies increased considerably when environmental variables were added. Accuracies were better for lower (50 m)-resolution environmental variables than for higher (5 and 10 m)-resolution environmental variables, but they were best for a combination of low- and high-resolution environmental variables (Table 1). Conversely, overall accuracies also increased when spring and autumn image data were added to all combinations of environmental variables. The increase was less when accuracies with environmental variables were already good, as in classifications with low resolution or low/high resolution. In winter, accuracies increased only when image data were added to the high-resolution environmental variables (Table 1).
Table 1. Overall accuracies (%) in the classification of 24 National Vegetation Classification classes and class ‘wood’ for areas photographed in three seasons and different combinations of predictor variables and resolutions
Means, ratios, SDs
No image data
If applicable, the resolution is given in brackets. Classifications with aerial image data for spring and autumn include flying date as an additional predictor variable. E, elevation; A, aspect; S, slope.
Image data only
E (10 m), A (10 m), S (5 m), soil
E (50 m), A (50 m), S (50 m), soil
E (10 and 50 m), A (10 and 50 m), S (5 and 50 m), soil
Image data only
E (10 m), A (10 m), S (5 m), soil
E (50 m), A (50 m), S (50 m), soil
E (10 and 50 m), A (10 and 50 m), S (5 and 50 m), soil
Image data only
E (10 m), A (10 m), S (5 m), soil
E (50 m), A (50 m), S (50 m), soil
E (10 and 50 m), A (10 and 50 m), S (5 and 50 m), soil
In all seasons, producer’s accuracies for class ‘wood’ were considerably improved through the addition of image data. Apart from winter, user’s accuracies for class ‘wood’ also increased with the inclusion of image data, considerably so in autumn. Producer’s and user’s accuracies for other classes also increased and for some classes decreased with the addition of image data, though not consistently across all seasons (see Table S1, Supporting Information).
Overall accuracies were between 87% and 92% using aspect, slope and elevation at the high and low resolution as well as soil, means, ratios, SDs and flying date (Table 1). Median producer’s class accuracy for this combination was 95% (58–100%) and median user’s class accuracy 92% (67–100%). Forty-nine per cent of producer’s and 29% of user’s class accuracies were higher than 95%. Twenty-five per cent of producer’s and 44% of user’s class accuracies were between 85% and 95%. Only 10% of producer’s and 1% of user’s class accuracies were below 70% (see Table S1, Supporting Information). Mantel tests did not detect spatial autocorrelation in the residuals (all three seasons: P >0·05).
We predicted 24 NVC classes and class ‘wood’ for the whole study area using aspect, slope and elevation at the high and low resolution as well as soil, means, ratios, SDs and flying date (Fig. 2). To get an indication of conformity between the predicted map and the mosaic polygons (mapping units in which several NVC communities had been recorded, see Vegetation Community Data), we recorded for each mosaic polygon whether the NVC classes recorded were also predicted in it. On average, 81% of recorded NVC classes were predicted in mosaic polygons.
In the predicted map, all 25 classes were spatially aggregated with a mean clumpiness index of 0·85 (0·80–0·92). Mean patch size per class varied from 2700 m2 for M10 (flush vegetation community) to 13 900 m2 for M20 (blanket bog). The commonest class was M19 (blanket bog), while CG2 (calcareous grassland) was the rarest class (see Table S2, Supporting Information).
Classification Accuracy of Training Pixels
We classified 24 NVC classes and a class ‘wood’ from a combination of environmental variables and aerial images with an overall accuracy between 87% and 92%. Our approach showed that even for relatively complex vegetation classifications, it is possible to achieve considerable accuracy using mostly remotely sensed data. The high accuracies are particularly remarkable given that some communities are difficult to distinguish even using trained field workers.
Average class accuracies were high with a median of 95% (58–100%) for producer’s and 92% (67–100%) for user’s class accuracy. Several classes produced comparatively low accuracies across several seasons. This may be attributable to a poorer mapping accuracy for these classes, particularly M6 (upland wetland), M19, M20 (blanket bogs), U4, U5 and U6 (upland acid grasslands). Especially in unenclosed land, these frequently tend to be recorded either as mosaics or as one class although they may contain patches of other classes.
Several predicted classes produced similar accuracies between seasons (e.g. H12, MG3, MG5), while others differed substantially between seasons (e.g. CG9, M19, U4). The contribution of aerial images to the classification is expected to vary with season (Lillesand, Kiefer & Chipman 2008). The fact that prediction accuracies also differed between seasons when aerial images were not included suggests that environmental variability and/or mapping accuracy for several classes varied between the areas covered by the three seasons (see Table S1, Supporting Information).
Resolution of Predictor Variables
The resolution of predictor variables influenced classification accuracy. For image data, using reflectance values and their ratios (means and ratios), the low resolution (5 m) produced better classification accuracies than the high resolution (0·25 m). Several studies have found that averaging high-resolution imagery to a lower resolution produced more accurate classification for some habitat classes as averaging over pixels reduced the heterogeneity in the reflectance values (Cushnie 1987; Wardley, Milton & Hill 1987). For environmental variables, a combination of low- (50 m) and high-resolution (5–10 m) predictors achieved best results, i.e. when variables described the neighbourhood and the local conditions. A location that is surrounded by dissimilar environments may be colonized by species less suited to its conditions, for example if these species are more likely to disperse to the location (Leathwick 1998; Miller, Franklin & Aspinall 2007). The average conditions in the surroundings of a location, i.e. environmental variables at a lower resolution, may thus be a good predictor of the communities most likely to succeed through intrinsic factors such as dispersal and competition, while high-resolution variables described the local conditions. The survey data were mapped at a comparatively coarse scale (1:5000 and 1:10 000), which might also explain or contribute to the good performance of the low-resolution environmental variables.
Comparison With Other Studies
High classification accuracies for vegetation maps were reported, for example, by Chapman et al. (2010), with c. 95% for seven land cover classes and four age classes of heather in the British uplands using aerial images, and by Sesnie et al. (2008), with 93% for 32 land cover types in Costa Rica using satellite images and environmental data. Often, however, lower classification accuracies were reported, for example 62% for eight peatland community classes in Canada using aerial images (Thomas et al. 2002), 54% and 80% for two seasons and 15 habitat classes in the British uplands using satellite images (Mehner et al. 2004), 69% for 17 alpine plant communities in Austria using aerial images and topographical variables at neighbourhood scales (50-m resolution and coarser) (Dirnböck et al. 2003), and 75% for 12 plant communities in California using environmental variables and aerial and satellite images (Dobrowski et al. 2008). For 27 NVC classes, Sanderson et al. (1995) reported 0·034–0·728 correlation between recorded NVC classes and a predicted habitat suitability index for the communities using environmental and management variables. Reid & Quarmby (2000) distinguished eight blanket bog classes in Scotland from satellite images and ancillary data, with 77% accuracy. A link between the derived classes and NVC communities was limited, as some NVC communities were common in several classes. Shanmugam et al. (2003) found for a coastal dune system in the United Kingdom that good classification accuracies (>90%) using aerial images could only be achieved when several NVC community classes were grouped into broader habitat types. Accuracies for mapping at the phase 1 level [a system for comparatively rapid classification of semi-natural habitats at a coarser level than NVC communities (England Field Unit Nature Conservancy Council 2003)] in Wales using image and ancillary data were reported for polygons in a subset of two of 42 areas and were 83% and 87% for a subset of 15 habitat classes (Lucas et al. 2011). The UK Land Cover Map 2000 predicts 16 broad habitat classes across the whole of the United Kingdom with a target accuracy of 90% and a resolution of 25 m. The classes can be further divided into subclasses and 71 variants although with a possibly reduced accuracy (Fuller et al. 2002; Centre for Ecology & Hydrology 2010).
The high classification accuracies achieved in this study (87–92%), the high number of classes differentiated (25), the level of detail of the classes (NVC classes rather than land cover classes) and the high resolution (5 m) indicate that this study is among the best performing to date.
NVC Communities in the Yorkshire Dales
The blanket bog communities M19 (11·7%) and M20 (9·7%) as well as semi-improved grassland MG6 (11·2%) were the most common predicted classes (see Table S2, Supporting Information). The greatest number of patches was recorded for class ‘wood’ (34342), reflecting the frequent occurrence of single trees and bushes in the study area.
Improvements and Outlook
We used existing data from a large-scale survey that were not collected for the purpose of vegetation modelling. Large-scale NVC surveys inherently contain mapping errors caused by the comparatively large mapping scale and time constraints. If NVC data were to be specifically recorded for vegetation modelling, more accurate mapping would be desirable, which may increase classification accuracies. The sampling design should ensure that sample patches are representative for the study area, so that maps for the whole area can be derived with confidence. Use of aerial or satellite data additional to the bands used in this study could probably increase classification accuracies. Plant species or communities can often be distinguished by their reflectance in the near-infrared bands (Morton 1986; Lillesand, Kiefer & Chipman 2008), which we did not have available. Images from digital cameras that avoid the colour reproduction differences between different batches and brands of film (Lillesand, Kiefer & Chipman 2008) and more sophisticated colour post-processing could probably avoid the colour differences between regions and flying dates encountered in this study. Further work should assess whether good classification accuracies can be achieved in habitats outside the study area with the proposed approach.
To our knowledge, this is the first study to achieve classification accuracies as high as 87–92% for a high number (24) of detailed NVC communities and an additional land cover class ‘wood’ with data from a geographically large extent. (On average, the training pixels were spread over 171 km2; median of minimum convex polygon). This methodology could save time and effort in the creation of NVC maps for large areas. Assuming that one person can map 1·25 km2 per day (Reid & Quarmby 2000), mapping NVC communities for the area shown in Fig. 2a would require 1304 person-days. In this study, training pixels from 2583 NVC polygons were used covering 111·7 km2. (However, it was beyond the scope of this study to assess whether the number and spread of training pixels used were optimal.) Assuming that only 0·7 km2 per day can be mapped owing to the higher accuracy desired for this application, the polygons could have been mapped in 160 person-days. This approach of combining the algorithm random forest with environmental and image variables at different spatial resolutions demonstrates the potential power of remotely sensed map making for ecological map making on very large spatial scales. In contrast to existing maps such as the UK Land Cover Map 2000 (Fuller et al. 2002), resulting maps would provide information at the community level rather than broad land cover types. Such maps would be valuable for applications including the monitoring of vegetation under climate change, the study of relationships between the occurrence of animal species and vegetation communities, habitat management or providing a planning base for land use decisions in environmentally valuable areas.
We thank Dan Chapman for introducing us to his image interpretation approach and stimulating discussions. We thank Peter W, two anonymous referees and the editors for constructive comments on earlier versions of this manuscript. The study was funded by the YDNPA. The OS MasterMap data were used under YDNPA OS Licence Number 1000237402007.