Assessing Urban Forest Canopy Cover Using Airborne or Satellite Imagery

Aerial Photograph Interpretation

Interpretation of aerial photographs at randomly placed locations in a study area has been used to determine the quantity of urban forest cover (Rowntree 1984; Nowak et al. 1996). The presence or absence of tree canopy cover at the specific point position is tallied for each of the sample points, and the proportion of sample points that fell on the tree canopy statistically represents the amount of urban forest canopy cover in the study area. A standard error of the sampling can also be computed to yield a bound on the canopy percent estimate. With the wide availability of digital orthophotographs and geographic information systems in many municipalities, aerial photograph interpretation is very simple to implement. It yields very good results for quantity, but provides limited information on the spatial distribution of the forest cover. Aerial photograph interpretation requires little technical remote-sensing knowledge beyond the ability to interpret tree canopies. Tools to aid in digital aerial photograph interpretation are available at www.fs.fed.us/ne/syracuse/Tools/tools.htm.

Several properties of the aerial imagery itself can influence the ease of interpretation and quality of results. Imagery acquired during the leaf-on season for trees will be more straightforward to interpret than leaf-off imagery. Leaf-off imagery may still be used for photograph interpretation, but more interpretation skill and time is required to infer canopy from the visible branch structure and/or shadows of the stem and branches on the ground surface. Resolution for digital images, or scale for traditional print photographs, should be such that individual trees can clearly be discerned. For digital imagery, this resolution would generally be approximately 1 m (3.3 ft) or smaller. The drawback of very high-resolution imagery is that the digital files will be quite large and it may be time-consuming to change views from one point to another. Color infrared (CIR) imagery is useful in differentiating various types of vegetation because CIR can accentuate the subtle differences between types of trees, between grass and trees, and between vegetation and nonvegetative surfaces. On all types of aerial and satellite images, trees overlap other ground surfaces such as buildings, roads, or parking lots (impervious surfaces) and bare soil, grass, or shrubs (permeable surfaces). Because the understory or ground cover cannot usually be seen, assessments of the full extent of surface features and canopy depth, useful for watershed analyses, for example, cannot be exactly determined from leaf-on aerial photographs or optical remotely sensed imagery.

Types of Digital Imagery

Several different types of digital aerial or satellite imagery have been used to assess urban forest canopy cover, including high-resolution digital aerial photographs, high-resolution satellite imagery, medium-resolution satellite imagery, and low-resolution satellite imagery.

High-resolution digital frame cameras have recently been used in place of traditional film aerial cameras (Myeong et al. 2001; Zhang 2001). The resulting imagery has the three portions of the spectrum (either red, green, and blue or near infrared [NIR], red, and green) separated into individual bands that can be used in further automated processing. These cameras generally produce images with submeter to meter spatial resolution, i.e., each image pixel represents an area on the ground 1 m × 1 m square or smaller. Emerge DSS (Applanix Corp., Richmond Hill, Ontario, Canada) and Leica ADS40 (Leica Geosystems AG, Heerbrugg, Switzerland) are examples of airborne sensor systems that produce this type of imagery.

High-resolution aerial and satellite imagery has been used to map urban forest cover in many cities, including Baltimore and Annapolis, Maryland, U.S. (Irani and Galvin 2002; Galvin et al. 2007); Syracuse, New York (Myeong et al. 2001); and New York City (www.oasisnyc.net/oasismap.htm). The most common high-resolution satellite imagery, from the Quickbird and IKONOS satellites, consists of four multispectral bands (NIR, red, green, and blue) with spatial resolution of 2.4 m (7.9 ft) (Quickbird) or 4 m (13.2 ft) (IKONOS) coupled with a 1 m (3.3 ft) (IKONOS) or 0.6 m (2 ft) (Quickbird) spatial resolution panchromatic band. The multispectral bands can be combined with the higher resolution grayscale panchromatic band to produce “pansharpened” multispectral imagery at the higher resolution. These sensors have enough spectral coverage to allow the imagery to be used successfully in automated classification algorithms. The spatial resolution is adequate to identify canopies of individual, open-grown trees.

Medium-resolution satellite imagery, with spatial resolution of tens to a few hundred meters, lends itself to analysis of forest cover of entire cities or regional urbanized areas. This spatial resolution depicts areas that can have several land cover types in an individual pixel. For example, in a residential area, tree canopies, building roofs, lawns, and paved streets could be represented within a single pixel. These “mixed pixels” are suited for specialized subpixel processing methods that are discussed subsequently. Enhanced Thematic Mapper (ETM) on the Landsat series of Earth imaging satellites is one of the most common examples of medium-resolution satellite imagery. Landsat’s ETM captures six multispectral bands, three in the visible part of the spectrum and the remainder in the NIR and midinfrared portion of the spectrum. Because living vegetation is highly reflective in the NIR, these bands facilitate the differentiation of vegetated surfaces from other land cover. Landsat imagery has been used to map urban tree cover (e.g., Wang 1988; Iverson and Cook 2000). In recent classifications from the National Land Cover Database (NLCD) 2001 (Homer et al. 2007), a tree canopy layer is derived from Landsat ETM imagery and is being used to assess urban forest cover across the United States as part of the USDA Forest Service’s Resources Planning Act (RPA) program.

Satellite imagery with a spatial resolution of several hundred meters and larger can be considered low resolution. Because of the complex mixing of land cover in developed areas and the large ground size of the pixels, low-resolution imagery is of limited usefulness for most urban forest cover mapping unless specialized subpixel analysis techniques are used. These techniques provide percent forest cover in each pixel and are described in the next section. Advanced Very-High Resolution Radiometer (AVHRR) imagery with 1.1 km (0.66 mi) spatial resolution has been used to assess urban forest cover in all the cities of the coterminous United States but had various limitations (Zhu 1994; Dwyer et al. 2000). The AVHRR sensor, primarily used for weather observations, collects data in one band in the green, one band in the reflected NIR, and three bands of the thermal parts of the spectrum so it is of marginal usefulness for general land cover mapping. The combination of low spatial and spectral resolutions of AVHRR produces highly variable results for urban forest cover mapping (Walton 2008).

Classification of Digital Imagery

Urban forest canopy cover can be mapped using the imagery described previously and one of many image classification techniques, including normalized difference vegetation index (NDVI), pixel-based, and object-oriented classification of high-resolution images and subpixel estimation of medium-resolution imagery.

Normalized Difference Vegetation Index. The NDVI is a very simple method used to accentuate vegetation from imagery containing reflectance in the red and the NIR portions of the spectrum. It is computed using the NIR and red (R) reflectance bands in the ratio as shown in Equation (1). The index produces high (brighter) values for pixels that contain more vegetation, but because it is a simple ratio, there can be other areas in an image such as sun glint from water surfaces that also produce high NDVI values. In addition, NDVI can be influenced by external factors (e.g., view angle, leaf orientation, and soil background; Campbell 2007) so that it does not produce a linear scale that is in direct proportion to the percentage of vegetation in the pixel.

(1)

Early remote-sensing analysis of urban forests using multispectral imagery such as that by the nonprofit group American Forests in the mid-1990s used the NDVI band ratio directly as a simple classification of urban forest cover. This approach is very limited because NDVI includes all vegetation and does not produce values that are linearly proportional to the amount of forest cover. Currently, NDVI is often used as an auxiliary input layer for other automated classification algorithms (e.g., Myeong et al. 2001).

High-Resolution Image Classification. Classification of high-resolution images poses difficulties for algorithms because the size of an individual pixel is much smaller than the size of the object to be classified resulting in pixels of different spectral characteristics making up the same object. Tree canopies, for example, when viewed on submeter high-resolution imagery, are made up of pixels representing various levels of reflectance from the vegetated surfaces as well as pixels of shaded areas. This disparity of the constituents of urban forest cover makes it difficult to classify tree canopies from their spectral signature alone. Several studies (Myeong et al. 2001; Zhang 2001; Irani and Galvin 2002) have used texture measures and, more recently, object-based methods (Walker and Briggs 2007; Zhou and Troy 2008) have been applied to high-resolution image classification of urban forest cover to overcome the difficulties presented by varying reflective properties among pixels representing one category.

One of the primary difficulties in classifying urban forest cover is the differentiation between tree canopies and grass surfaces. At an individual pixel level, tree canopies and grass surfaces generally have a very similar spectral response. However, tree canopies generally contain areas of bright, higher reflectance values from vegetation surfaces and areas of darker, low reflectance values from the shadowy regions of the tree canopy. This effect creates a more coarse texture on the image of a tree canopy when compared with other vegetated surfaces such as grassy lawns or athletic fields, which generally have a smoother, finer texture. Algorithms designed to identify regions of coarse texture have been shown to be very useful in differentiating tree and shrub canopies from other vegetation surfaces (Myeong et al. 2001; Zhang 2001).

Another method to handle highly variable pixel characteristics of urban forest canopy is through the use of object-based classification. Images are segmented into clusters of similar pixels by considering any number of pixel-based metrics such as spectral or textural values. These segments, or objects, are then combined to form the final classification categories. Considering objects rather than individual pixels yield results that mimic human classification. Object-based image analysis has been used by Walker and Briggs (2007), O’Neil-Dunne (2007), and Galvin et al. (2007). Zhou and Troy (2008) classified urban land cover at the parcel level using object-oriented techniques. Object-based analyses can integrate many different types of input data. Airborne LIDAR (laser ranging) data have been used to incorporate vegetation height information into the classification process (O’Neil-Dunne 2007).

Subpixel Estimation. Traditionally, techniques using medium-resolution imagery have used “whole-pixel” classifications in which each pixel is designated as either “forested” or “not forested” with some threshold being used to determine the cutoff between the two classes. Because of the heterogeneous nature and number of mixed pixels in urban areas, whole-pixel classifications tend to misrepresent the amount and spatial distribution of urban forest cover. For subpixel urban forest cover, the goal is to estimate the percent tree canopy cover for each pixel as a real number between zero and 100. Zhu (1994) developed a subpixel forest density map using a regression procedure from 1.1 km (0.66 mi) AVHRR multispectral imagery, which was later used to assess urban forest cover by Dwyer et al. (2000). The nonprofit group American Forests (2007) has used the ERDAS Imagine Subpixel Classifier (Applied Analysis, Inc., www.discover-aai.com/software/products/IMAGINE_Subpixel_Classifier.htm) to map and assess change of urban forest canopy. Small (2001) and Small and Lu (2006) have used spectral mixture modeling to estimate subpixel vegetation abundance in the New York City region. Recently, with the subpixel tree canopy layer from the NLCD 2001 (Homer et al. 2007), urban forest cover and accuracy assessments are being conducted using this 30 m, nationwide resource by the USDA Forest Service, Syracuse, New York. The NLCD 2001 tree canopy estimate was generated using the Cubist (RuleQuest Research, www.rulequest.com) rule-based regression software (Huang et al. 2001).

Subpixel urban forest cover maps generally represent the amount and spatial distribution of urban forest canopy cover better than whole-pixel classification techniques for medium-resolution imagery.

Classification Accuracy Assessment. For all methods of digital image classification, an accuracy assessment of the map should be conducted to determine how accurately the map categories (e.g., tree cover) were classified. One relatively simple way to determine map classification accuracy is to randomly sample points throughout the classified area and ground-truth the classification at those points either through site visits or aerial photograph interpretation. The map classification at each point is then compared with the ground-truth estimate to determine map classification accuracy (Table 1). Typically, overall map accuracies within the 80% to 90% accuracy range can be obtained, but accuracy varies among individual classification categories. As a result of the varying classification techniques and imagery products, understanding map classification accuracy is essential to understanding how useful the map will be for planning, management, and analysis purposes.

Assessing Urban Forest Canopy Change

Many of these map products discussed here can be used as the basis of urban forest canopy change studies. These assessments can vividly demonstrate the effects of urbanization on the forest canopy resource and bolster public opinion with regard to local land development policies. Understanding how urban forest canopy cover is changing can be valuable information for the forest managers. Several groups have used imagery as the means to assess canopy cover change, including the nonprofit group American Forests (2007) in several U.S. cities and Poracsky and Lackner (2004) in Portland, Oregon. Two important factors in determining canopy cover change are 1) assessing change over a long enough time period for change to be evident; and 2) having a consistent methodology when developing each individual canopy map that is compared (Walton 2008). A consistent methodology will include using similar imagery and analysis methods when developing both products used in the comparison.