Impervious surface extraction in urban areas from high spatial resolution imagery using linear spectral unmixing

Remote Sensing Applications: Society and Environment 1 (2015) 61–71

Contents lists available at ScienceDirect

Remote Sensing Applications: Society and Environment journal homepage: www.elsevier.com/locate/rsase

Impervious surface extraction in urban areas from high spatial resolution imagery using linear spectral unmixing Jian Yang a,n, Peijun Li b a b

Department of Geography, University of Toronto, 100 St. George Street, Toronto, ON, Canada M5S 3G3 Institute of Remote Sensing and GIS, School of Earth and Space Sciences, Peking University, Beijing 100871, PR China

a r t i c l e i n f o

abstract

Article history: Received 17 March 2015 Received in revised form 6 June 2015 Accepted 10 June 2015 Available online 28 July 2015

Impervious surface, as an important indicator of urbanization assessment, plays a significant role in analyzing the climate, environment and hydrologic cycle in urban areas. Impervious surface extraction in urban areas from satellite imagery attracts growing attention in many applications. Recently, the increasing availability of high spatial resolution satellite imagery provides new opportunities for impervious surface extraction at a fine scale. However, impervious surface like asphalt roads, parking lots, and sidewalks is often obscured by tree canopies, which can remarkably underestimate impervious surface area in urban areas. In order to overcome this problem, this study adopts linear spectral unmixing to detect impervious surface information through tree canopies, and further incorporates with object-based classification to mitigate the negative effects of tree canopy obscurity when extracting impervious surface from high spatial resolution imagery. The performance of the proposed impervious surface extraction method is investigated by a subset of QuickBird imagery in Beijing urban areas. Results demonstrate that the proposed method effectively reduces impervious surface underestimation in urban areas by 11.20%, and more accurate impervious surface extraction and mapping can assist government policy makers for timely monitoring of urban hydrological environment. & 2015 Elsevier B.V. All rights reserved.

Keywords: Impervious surface extraction Tree canopy obscurity Linear spectral unmixing SESMA MESMA High spatial resolution

1. Introduction With rapid urbanization in the past decades, more and more land surface is covered by urban built-up areas. One of the most obvious physical evidences for urban growth is an increasing area of impervious surface (Jensen and Cowen, 1999). Impervious surface where water cannot infiltrate through the ground, including building rooftops, asphalt roads, highways, parking lots, and sidewalks, directly affects the amount of runoff to streams and lakes and non-point source pollution even water quality (Dougherty et al., 2004). Therefore, impervious surface distribution becomes an important indicator for monitoring urban hydrological n

Corresponding author. Tel.: þ 1 416 978 3375. E-mail address: [email protected] (J. Yang).

http://dx.doi.org/10.1016/j.rsase.2015.06.005 2352-9385/& 2015 Elsevier B.V. All rights reserved.

environment (Choi and Ball, 2002; Shuster et al., 2005; Zhou et al., 2010). In recent years, new satellite-based sensors promise to transform traditional impervious surface mapping that has either relied on expensive ground-based measurements or less accurate interpretation of aerial photography. Particularly, the increasing availability of high spatial resolution satellite imagery, such as IKONOS, QuickBird, GeoEye, and WorldView, has offered new opportunities for impervious surface extraction in urban areas at a fine scale (i.e., under 5 m) (Cablk and Minor, 2003; Goetz et al., 2003; Lu et al., 2011; Lu and Weng, 2009; Yuan and Bauer, 2006). Besides, more and more studies prefer the object-based image analysis in contrast with the pixel-based methods because the object-based methods perform better with less “salt and pepper” noises in high spatial resolution imagery (Laliberte et al., 2004; Zhou and Troy, 2008). Since land covers in

62

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

shadow areas show different spectral features from those in non-shadow areas, it is more preferable to separately extract impervious surface from shadow and non-shadow areas, hereafter namely hierarchical impervious surface extraction (De Roeck et al., 2009; Li et al., 2011). Impervious surface including asphalt roads, parking lots, and sidewalks is often obscured by tree canopies, which can remarkably underestimate impervious surface area in urban areas (Van der Linden and Hostert, 2009). Nevertheless, tree canopies, especially along the street, are not completely occlusive so some impervious surface (e.g., asphalt roads) information can be detected through tree canopies in high spatial resolution imagery. The pixels containing impervious surface information are very likely mixed with the spectral features of tree canopies. Conventionally, linear spectral unmixing is commonly used for impervious surface mapping in medium-low resolution or hyperspectral imagery (Deng et al., 2012; Van de Voorde et al., 2009; Weng et al., 2009). However, both Small (2003) and Yang et al. (2014) found that the inversion of a simple three-component linear mixture model is able to produce stable, consistent estimates of endmember fractions for each pixel even in high spatial resolution imagery. Despite not so many mixed pixels in high spatial resolution imagery, the vegetation-high albedo-low albedo model can provide a widely applicable physical characterization of impervious surface in urban areas. Thus, this study chose to adopt linear spectral unmixing to detect impervious surface information through tree canopies from high spatial resolution imagery.

2. Study area and experimental data Beijing urban area, China was selected as the study area. Land cover types in the study area include vegetation cover such as trees and grassland, and impervious surfaces such as sidewalks and asphalt roadways. QuickBird imagery of Beijing urban area, acquired on Sep. 30 of 2003 (off-nadir viewing angle: 9.3 degrees) was used in the study. The QuickBird imagery contains four multispectral bands with the spatial resolution of 2.44 m (blue, green, red, and near infrared band) and a panchromatic band with the spatial resolution of 0.61 m. Moreover, the multispectral and panchromatic images were fused to produce a pan-sharpened multispectral image with the pixel size of 0.61 m as the reference imagery. Specifically, the image fusion was carried out using the Gram–Schmidt procedure (Laben and Brower, 2000) implemented in the ENVI software package. A subset of the multispectral image of 918  923 pixels was adopted in this experiment (Fig. 1), corresponding with the reference imagery of 3672  3692 pixels.

3. Methods The proposed impervious surface extraction method included two main parts, that is linear spectral unmixing and object-based classification. In terms of linear spectral unmixing, both Simple Endmember Spectral Mixture

Analysis (SESMA) and Multiple Endmember Spectral Mixture Analysis (MESMA) models were implemented and quantitatively compared for impervious surface information detection through tree canopies. Hierarchical objectbased classification was then conducted for impervious surface extraction in shadow and non-shadow areas, respectively. Further, the final impervious surface map was derived by the appropriate extension of the detected impervious surface information through tree canopies based on the object-based classification result. At last, the mapping accuracy of impervious surface was quantified by the reference impervious surface map. Detailed procedures are demonstrated as Fig. 2. 3.1. Linear spectral unmixing 3.1.1. Shadow restoration Since most of impervious asphalt roads in urban areas are absorptive at optical wavelengths (Herold and Roberts, 2005; Herold et al., 2004), shadow restoration is of great necessity prior to linear spectral unmixing because it is able to reduce spectral confusion between shadow areas and low albedo land covers in non-shadow areas (e.g., asphalt roads). In lieu of the commonly applied histogram threshold method (Chen et al., 2007; Dare, 2005; Zhou et al., 2009), this study made use of object-based classification to identify the shadow areas over the entire imagery, which was also the first step of the following objectbased classification. A linear-correlation correction approach was thereafter utilized for shadow restoration. Several previous studies proved that it outperformed other shadow restoration techniques, such as Gamma correction and histogram matching (Dare, 2005; Sarabandi et al., 2004; Yang et al., 2015). The linear-correlation correction method hypothesizes that the signals recoded in shadow areas still provide enough useful information for restoration although they are relatively weak. Specifically, it assumes a linear relation between the digital number (DN) values of shadow areas and those of non-shadow areas, which can be expressed as (Chen et al., 2007; Sarabandi et al., 2004; Zhou et al., 2009)

y=

Sy (x − xm ) + ym Sx

(1)

where x and y are the original and output DN values of shadow pixels, Sx and Sy are the standard deviation of shadow and non-shadow areas while xm and ym are the mean values of shadow and non-shadow areas. 3.1.2. Endmember selection Since vegetation and impervious surface are two of the primary land covers in urban areas, the endmember classes of vegetation, high albedo, and low albedo impervious surface can provide a useful physical description of urban areas (Small, 2003). As such, this study also used these three endmember classes for linear spectral unmixing. In the MESMA model, too large number of endmembers in one class may result in significant challenges in interpretation and computation. So several endmember

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

optimization strategies have been proposed in recent years, such as Count-based Endmember Selection (CoB) (Roberts et al., 2003), Endmember Average RMSE (EAR) (Dennison and Roberts, 2003), and Minimum Average Spectral Angle (MASA) (Dennison et al., 2004). Specifically, this study employed MASA to identify the most representative endmembers within a specific class. MASA

63

calculates the average spectral angle between a reference spectrum (candidate model) and all other spectra within the same class, and the best endmembers are selected as the ones that produce the lowest average spectral angles (Dennison et al., 2004). Meanwhile, the mean value of the optimized representatives in each class was calculated as the endmember for the SESMA model. 3.1.3. Linear spectral unmixing models Linear spectral mixture model yields abundance estimate of each endmember class for an unmixed pixel, mathematically denoted as n

ρi =

∑ f j ρji

+ εi

j=1

Fig. 1. QuickBird imagery of this study area with a band combination of red, near infrared, and green as R, G, and B. It is under WGS84 UTM coordinate system. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

(2)

where ρi is the DN value of the unmixed pixel in spectral band i, f j is the fraction of the endmember class j for this pixel, ρji is the DN value of its corresponding endmember class j in spectral band i, and εi is the residual error. When the number of spectral bands is larger than that of endmember classes, Eq. (2) is called over-determined equation. In this case, f j is usually calculated through leastsquares solution in order to minimize the residual error. Additionally, the SESMA model also requires the fraction of each endmember class range from zero to one meanwhile the sum of all the fractions within any pixel is equal to one. Therefore, this study implemented the constrained leastsquares solution for the SESMA model. Although the SESMA model is a powerful approach for linear spectral unmixing, a single endmember cannot

Fig. 2. Flowchart of impervious surface extraction procedures proposed by this study.

64

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

account for considerable spectral variation within a class. MESMA extends SESMA by allowing a number of endmembers to vary on a pixel basis (Roberts et al., 1998). With regard to any pixel, MESMA selects an endmember from each class and seeks for the best combination of endmember spectra from all possible combinations. From this perspective, MESMA overcomes the limitations of SESMA by requiring an optimal model to meet the selection criteria include fractional constraints (minimum and maximum fraction constrains), RMSE constraint and residual constraints while testing multiple models for each pixel (Franke et al., 2009; Michishita et al., 2012; Powell et al., 2007; Rosso et al., 2005). Thus, this study also utilized MESMA implemented with VIPER Tools, an ENVI plugin developed by Roberts et al. (2007). 3.1.4. Spectral unmixing accuracy assessment Ideally, ground-based field measurement is demanded for accuracy assessment of linear spectral unmixing. Unlike real land covers, however, it was difficult to ground-based measure the samples of high albedo and low albedo impervious surface in a consistent way with what are shown in the imagery. Differentiation of high albedo and low albedo impervious surface is quite dependent upon not only solar angle and intensity but also viewing angle of satellite sensor. In addition, land covers in Beijing where urban renewal processed at a rapid speed have greatly changed over the past decade, and the QuickBird imagery used in this experiment was not the latest data. So this study adopted the classification result from the pan-sharpened multispectral imagery as the reference to evaluate the unmixing performance of the SESMA and MESMA models. Corresponding to the selected endmember classes, the pan-sharpened multispectral imagery (shadow restored) was classified into the same three classes including vegetation, high albedo, and low albedo impervious surface. Considering the spatial resolution of the original multispectral imagery (2.44 m) is four times as that of the pansharpened imagery (0.61 m), each pixel in the original multispectral imagery is equivalent to 4  4 pixels in the pan-sharpened imagery. For an unmixed pixel, the abundance of each endmember class was computed based on the classified reference imagery. In this study, we randomly selected 1000 pixels over the entire imagery and did a linear regression analysis between the unmixed fraction and the calculated abundance for each endmember class. Consequently, the model with a higher R2 would be used to detect impervious surface information through tree canopies. 3.2. Object-based classification 3.2.1. Hierarchical impervious surface mapping As this study focused on impervious surface extraction by eliminating tree canopy obscurity, grassland should be differentiated from the vegetation class. Thus, this study area was classified into three land covers including tree, grassland, and impervious surface. As mentioned previously, the reduction or total loss of spectral information in shadow areas may cause some misclassification between

shaded vegetation and unshaded dark impervious building rooftops (Lu and Weng, 2009), particular in high spatial resolution imagery where shadows are extensively distributed (Li et al., 2011). In this study, the shadow and nonshadows areas were first separated as shadow detection, and image classification was then implemented in the shadow and non-shadow areas, respectively. Furthermore, the classification results of the shadow and non-shadow areas were merged together. Specifically, one of the stateof-the-art classifier, the support vector machines (SVMs) (Shawe-Taylor and Cristianini, 2000), was employed for hierarchical image classification. For hierarchical object-based classification, multi-scale image segmentation is required. We adopted multiresolution segmentation algorithm, which is implemented in eCognition Developer 8.7. The segmentation procedure starts with one-pixel objects and merges similar neighboring objects together in subsequent steps until a heterogeneity threshold, set by a “scale parameter”, is reached (Benz et al., 2004). In this study, appropriate scale parameters were selected through trial-and-error exploration. In particular, one coarser scale parameter was needed to separate the shadow and non-shadow areas while two finer scale parameters were used for image classification in the shadow and non-shadow areas, respectively. Moreover, classification in the shadow areas demanded a relatively fine scale parameter because land covers in the shadow areas are usually smaller than those in the non-shadow areas. 3.2.2. Classification accuracy assessment Error matrix, also known as confusion matrix, was commonly used for quantitative evaluation of classification accuracy. An error matrix is a set of rows and columns that express the number of units (i.e. pixels or image objects) assigned to each class relative to the actual class as verified by reference data (Congalton, 1991). Accuracy parameters derived from an error matrix include producer’s accuracy (PA), user's accuracy (UA), overall accuracy (OA), and kappa coefficient (KC). Compared to the other parameters, an increasing number of studies criticized the effectiveness of KC (Foody, 2002; Liu et al., 2007; Pontius and Millones, 2011) because it actually does not serve a useful role in classification accuracy assessment (Olofsson et al., 2014, 2013). Thus, we chose to abandon KC for accuracy assessment, and followed the ideas recommended by Olofsson et al. (2014) to calculate the estimated PA, UA, and OA with standard errors (795% confidence interval). 3.3. Extension of impervious surface information Impervious surface information through tree canopies was identified by linear spectral unmixing. If a piece of tree canopy contains the detected impervious surface information, the land covers under this piece of tree canopy are very likely impervious surface. In this case, the pixels within this piece of tree canopy are marked as impervious surface, which is the extension of the detected impervious surface information. Specifically, the extent of a piece of tree canopy was determined by the segments that belong to the tree class.

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

For further comparison, this study employed the buffer analysis to obtain the main roads over the study area based on the road map of Beijing, and overlapped them on the classification result in order to eliminate street tree canopy obscurity. It was considered as the reference impervious surface map to validate the proposed impervious surface extraction method.

4. Results and discussion 4.1. Linear spectral unmixing Based on the mean and standard deviation values of the shadow and non-shadow areas (Table 1), the DN values of land covers in the shadow areas were restored for each spectral band using Eq. (1). A proportion of study area before and after shadow restoration is depicted in Fig. 3. The spectral features of the shaded vegetation (red ellipses in Fig. 3a) are well recovered (Fig. 3b) to strengthen the spectral contrast to low albedo impervious surface. A set of endmembers, consisting of 32, 42 and 47 pixels, were initially selected for each class of vegetation, high albedo, and low albedo impervious surface. The endmembers were Table 1 Mean and standard deviation (Std. dev.) values of the shadow and nonshadow areas for all the spectral bands (i.e., near infrared, red, green, and blue). Spectral band

Mean (nonshadow)

Std. dev. (nonshadow)

Mean Std. dev. (shadow) (shadow)

Near infrared Red Green Blue

336.7 266.5 416.1 308.5

65.2 54.8 50.1 24.6

233.9 231.7 386.5 297.5

35.3 25.4 24.2 12.1

65

further optimized to 15, 11, and 18 pixels with the lower MASA values as the representatives of vegetation, high albedo, and low albedo impervious surface, respectively. As a result, there were 2970 models in total for the MESMA model. Meanwhile, the mean values of the optimized endmembers are calculated for the SESMA model. The linear spectral unmixing results based on the SESMA and MESMA models are demonstrated in Fig. 4. For the SESMA model, the unmixed fraction maps of vegetation, high albedo, and low albedo impervious surface are shown as Fig. 4a–c, respectively. By contrast, the unmixed fraction maps derived by the MESMA model are depicted as Fig. 4d–f, respectively. It is worth noting that all the unmixed fractions are between zero to one, and a brighter pixel in the unmixed fraction map indicates a larger fraction of the corresponding class. The linear relations between the unmixed fractions (linear spectral unmixing) and the calculated abundances (reference imagery classification) are demonstrated for vegetation, high albedo, and low albedo impervious surface in Fig. 5. For each endmember class, the MESMA model outperforms the SESMA model due to the higher R2 (vegetation: 0.7239 vs. 0.6395; high albedo impervious surface: 0.7468 vs. 0.619; low albedo impervious surface: 0.6579 vs. 0.6083). In addition, all the slopes using the MESMA model are much closer to one than those using the SESMA model, which the MESMA results match the real land covers better. Thus, this study chose the linear spectral unmixing results derived by the MESMA model to detect the impervious surface information through tree canopies. The red ellipses in Fig. 6 demonstrate the detected impervious surface information from low albedo impervious surface fraction map using the MESMA model. Furthermore, the pixels with the unmixed fraction of high albedo impervious surface greater than 0.3, or the unmixed fraction of low albedo impervious surface greater

Fig. 3. A proportion of study area before (a) and after (b) shadow restoration. (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

66

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

Fig. 4. Unmixed fraction maps derived by the SESMA model of vegetation (a), high albedo (b), and low albedo impervious surface (c); unmixed fraction maps derived by the MESMA model of vegetation (d), high albedo (e), and low albedo impervious surface (f).

than 0.5, are considered as impervious surface information for the further extension. 4.2. Object-based classification As mentioned above, multiple scales of segmentation results were obtained for hierarchical object-based classification. After a number of experiments and comparison, this study selected 30 as the scale parameter for separating the shadow and non-shadow areas, thereafter 10 and 20 as the scale parameters for classification in the shadow and non-shadow areas, respectively. The SVM classifier with the radial basis function (RBF) was employed for hierarchical classification. First of all, the entire imagery was classified into the shadow and non-shadow areas. Specifically, a set of training samples including shadow areas (50 polygons) and non-shadow areas (50 polygons) were randomly selected. Then, the shadow and non-shadow areas were reclassified into tree, grassland, and impervious surface, respectively. In this step, two sets of training samples were randomly selected (for the shadow areas: 10 polygons of tree, 8 polygons of grassland, and 15 polygons of impervious surface; for the non-shadow areas: 20 polygons of tree, 20 polygons of grassland, and 20 polygons of impervious surface). Finally, the classification results in the shadow and non-shadow areas were combined to the entire imagery, as shown in Fig. 7.

In order to evaluate the hierarchical object-based classification accuracy, we randomly selected a set of test samples (i.e., 27 polygons for tree, 17 polygons of grassland, and 30 polygons for impervious surface), moreover constructed the error matrix of pixels counts (Table 2). As a result, the estimated UA (795% confidence interval) is 0.87270.013 for tree, 0.72070.022 for grassland, and 0.93670.008 for impervious surface. The estimated PA is 0.79470.017 for tree, 0.72070.033 for grassland, and 0.971 70.004 for impervious surface. The estimated OA is 0.90770.007. It is obvious that the classification quality was negatively impacted by the misclassification between tree and grassland despite the acceptable overall classification accuracy. 4.3. Extension of impervious surface information By the extension of the detected impervious surface information, the final impervious surface map is shown in Fig. 8. Specifically, the tree canopies to be used for extension were the segments classified as tree at scale parameter of 30. After eliminating the effects of tree canopy obscurity, the proportion of impervious surface in this study area increased from 68.91% to 76.63%, by 11.20%. Therefore, it could be concluded that that tree canopy obscurity remarkably overestimated the area of pervious

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

67

Fig. 5. Unmixed fractions derived by the SESMA model (horizontal axis) vs. calculated abundances (vertical axis) of vegetation (a), high albedo (b), and low albedo impervious surface (c); unmixed fractions derived by the MESMA model (horizontal axis) vs. calculated abundances (vertical axis) of vegetation (d), high albedo (e), and low albedo impervious surface (f).

68

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

Fig. 6. Impervious surface information through tree canopies shown in original multispectral imagery (a) and detected by linear spectral unmixing (b). (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

Fig. 7. Hierarchical object-based classification result (green: tree; blue: grassland; yellow: impervious surface). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 8. Final impervious surface map using the proposed method (white: impervious surface; black: pervious surface).

Table 3 Width of Road in Beijing. Table 2 Error matrix of pixel counts constructed from the accuracy assessment. Classified categories are the rows while the reference categories are the columns. Impervious Grassland Tree

Tree Grassland Impervious Total

179 43 3100 3322

143 1201 12 1356

2198 424 200 2822

Total

2520 1668 3312 7500

Map area (ha)

Wi

127 30 348 504

0.251 0.060 0.689 1

surface whereas underestimated the area of impervious surface in urban areas. On the other hand, the reference impervious surface map was produced by overlapping the main roads on the hierarchical object-based classification result, which was only to eliminate tree canopy obscurity along the main

Road types in Beijing

1st Road

2nd Road

3rd Road

4th Road

Road width (m)

50–70

40–60

30–50

10–30

roads. The roads in Beijing are mainly divided into: 1st Road, 2nd Road, 3rd Road and 4th Road. The road frames were delineated based on the road map in this study area. According to the Width of Road in Beijing (Table 3), the widths of 1st Road, 2nd Road, 3rd Road and 4th Road are set as 60 m, 50 m, 40 m, and 20 m, respectively. So we created 30 m, 25 m, 20 m, and 10 m buffer for each level of road frames to obtain the main roads of this study area (Fig. 9a). The reference impervious surface map is depicted as Fig. 9b. When using the reference impervious surface map for accuracy assessment of the final impervious surface map produced by the proposed method, the UA is 0.928 for impervious surface and 0.788 for pervious surface, while

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

69

Fig. 9. Main roads of this study area (a, yellow: main roads) and reference impervious surface map (b, white: impervious surface; black: pervious surface). (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 10. Illustration of impervious surface maps (a: original multispectral imagery; b: impervious surface map derived by the preliminary classification result; c: impervious surface map derived by the proposed method; d: reference impervious surface map). (For interpretation of the references to color in this figure, the reader is referred to the web version of this article.)

70

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

the PA is 0.935 for impervious surface and 0.769 for pervious surface. The OA is 0.895. A comparison of impervious surface maps is demonstrated in Fig. 10. For the asphalt roads obscured by street tree canopies (e.g., the red ellipses in Fig. 10), the impervious surface map derived by the proposed method (Fig. 10c) successfully eliminated the effects of tree canopy obscurity compared to the preliminary classification result (Fig. 10b). Meanwhile, the purple ellipse, which is actually an impervious asphalt road, however remains the same (Fig. 10b and c) because no impervious surface information could be detected though this piece of tree canopy. That is the reason why the UA of pervious surface is not very high. Moreover, the reference impervious surface map was produced without considering the pathways in residential and greening land (e.g., blue ellipse in Fig. 10), whereas the proposed method well solved this problem. In this case, the PA of pervious surface could also remarkably decrease when regarding this reference impervious surface map as the ground truth. Further, we must admit that impervious surface mapped using the proposed method might be a bit overestimated. For instance, land covers under tree canopies close to grassland (yellow ellipse in Fig. 10) are very likely grassland but they were updated to impervious surface. From this point of view, the proposed impervious surface extraction method has much potential to improve, especially in terms of identifying the appropriate size of tree canopy to be used for extension.

5. Conclusions and future work In this study, we proposed a new impervious surface extraction method to eliminate the negative effects of tree canopy obscurity. Linear spectral unmixing was adopted to detect the impervious surface information through tree canopies. The experimental results showed that the proposed method effectively reduced impervious surface underestimation in urban areas and well suited the real land covers. It is worth noting that the success of the proposed method may rely on both tree canopy openness and off-nadir viewing angle. With not so occlusive tree crowns and relatively vertical satellite angle, impervious surface information through tree canopy could be detected by linear spectral unmixing. Although the proposed method was effective to improve the accuracy of impervious surface extraction in urban areas, some further work should be still considered. Most importantly, more accurate and reliable linear spectral unmixing using high spatial resolution imagery are strongly required for this study. Additionally, other problems including distinguishing tree from other vegetation species like grassland even shrub which may exist in some urban areas (Ardila et al., 2012, 2011), and identifying the appropriate tree canopy to be used for extension also remain to be solved in the future work.

Acknowledgments The authors wish to acknowledge the Editor-in-Chief, Dr. Molly Brown, and two anonymous reviewers for their comments to improve the quality of this manuscript.

References Ardila, J.P., Bijker, W., Tolpekin, V.A., Stein, A., 2012. Context-sensitive extraction of tree crown objects in urban areas using VHR satellite images. Int. J. Appl. Earth Obs. Geoinf. 15, 57–69. Ardila, J.P., Tolpekin, V.A., Bijker, W., Stein, A., 2011. Markov-random-fieldbased super-resolution mapping for identification of urban trees in VHR images. ISPRS J. Photogramm. Remote Sens. 66 (6), 762–775. Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS J. Photogramm. Remote Sens. 58 (3–4), 239–258. Cablk, M., Minor, T., 2003. Detecting and discriminating impervious cover with high-resolution IKONOS data using principal component analysis and morphological operators. Int. J. Remote Sens. 24 (23), 4627–4645. Chen, Y., Wen, D., Jing, L., Shi, P., 2007. Shadow information recovery in urban areas from very high resolution satellite imagery. Int. J. Remote Sens. 28 (15), 3249–3254. Choi, K., Ball, J.E., 2002. Parameter estimation for urban runoff modelling. Urban Water 4 (1), 31–41. Congalton, R.G., 1991. A review of assessing the accuracy of classifications of remotely sensed data. Remote Sens. Environ. 37 (1), 35–46. Dare, P.M., 2005. Shadow analysis in high-resolution satellite imagery of urban areas. Photogramm. Eng. Remote Sens. 71 (2), 169–177. De Roeck, T., Van de Voorde, T., Canters, F., 2009. Full hierarchic versus non-hierarchic classification approaches for mapping sealed surfaces at the rural-urban fringe using high-resolution satellite data. Sensors 9 (1), 22–45. Deng, Y., Fan, F., Chen, R., 2012. Extraction and analysis of impervious surfaces based on a spectral un-mixing method using Pearl River Delta of China Landsat TM/ETMþ imagery from 1998 to 2008. Sensors 12, 1846–1862. Dennison, P.E., Halligan, K.Q., Roberts, D.A., 2004. A comparison of error metrics and constraints for multiple endmember spectral mixture analysis and spectral angle mapper. Remote Sens. Environ. 93 (3), 359–367. Dennison, P.E., Roberts, D.A., 2003. Endmember selection for multiple endmember spectral mixture analysis using endmember average RMSE. Remote Sens. Environ. 87 (2–3), 123–135. Dougherty, M., Dymond, R.L., Goetz, S.J., Jantz, C.A., Goulet, N., 2004. Evaluation of impervious surface estimates in a rapidly urbanizing watershed. Photogramm. Eng. Remote Sens. 70 (11), 1275–1284. Foody, G.M., 2002. Status of land cover classification accuracy assessment. Remote Sens. Environ. 80 (1), 185–201. Franke, J., Roberts, D.A., Halligan, K., Menz, G., 2009. Hierarchical Multiple Endmember Spectral Mixture Analysis (MESMA) of hyperspectral imagery for urban environments. Remote Sens. Environ. 113 (8), 1712–1723. Goetz, S.J., Wright, R.K., Smith, A.J., Zinecker, E., Schaub, E., 2003. IKONOS imagery for resource management: Tree cover, impervious surfaces, and riparian buffer analyses in the mid-Atlantic region. Remote Sens. Environ. 88 (1–2), 195–208. Herold, M., Roberts, D., 2005. Spectral characteristics of asphalt road aging and deterioration: implications for remote-sensing applications. Appl. Opt. 44 (20), 4327–4334. Herold, M., Roberts, D.A., Gardner, M.E., Dennison, P.E., 2004. Spectrometry for urban area remote sensing—development and analysis of a spectral library from 350 to 2400 nm. Remote Sens. Environ. 91 (3–4), 304–319. Jensen, J.R., Cowen, D.C., 1999. Remote sensing of urban/suburban infrastructure and socio-economic attributes. Photogramm. Eng. Remote Sens. 65, 611–622. Laben, C.A., Brower, B.V., 2000. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening. Google Patents. Laliberte, A.S., Rango, A., Havstad, K.M., Paris, J.F., Beck, R.F., McNeely, R., Gonzalez, A.L., 2004. Object-oriented image analysis for mapping shrub encroachment from 1937 to 2003 in southern New Mexico. Remote Sens. Environ. 93 (1), 198–210.

J. Yang, P. Li / Remote Sensing Applications: Society and Environment 1 (2015) 61–71

Li, P., Guo, J., Song, B., Xiao, X., 2011. A multilevel hierarchical image segmentation method for urban impervious surface mapping using very high resolution imagery. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 4 (1), 1–14. Liu, C., Frazier, P., Kumar, L., 2007. Comparative assessment of the measures of thematic classification accuracy. Remote Sens. Environ. 107 (4), 606–616. Lu, D., Hetrick, S., Moran, E., 2011. Impervious surface mapping with QuickBird imagery. Int. J. Remote Sens. 32 (9), 2519–2533. Lu, D., Weng, Q., 2009. Extraction of urban impervious surfaces from an IKONOS image. Int. J. Remote Sens. 30 (5), 1297–1311. Michishita, R., Jiang, Z., Xu, B., 2012. Monitoring two decades of urbanization in the Poyang Lake area, China through spectral unmixing. Remote Sen. Environ. 117, 3–18. Olofsson, P., Foody, G.M., Herold, M., Stehman, S.V., Woodcock, C.E., Wulder, M.A., 2014. Good practices for estimating area and assessing accuracy of land change. Remote Sens. Environ. 148, 42–57. Olofsson, P., Foody, G.M., Stehman, S.V., Woodcock, C.E., 2013. Making better use of accuracy data in land change studies: estimating accuracy and area and quantifying uncertainty using stratified estimation. Remote Sens. Environ. 129, 122–131. Pontius Jr, R.G., Millones, M., 2011. Death to Kappa: birth of quantity disagreement and allocation disagreement for accuracy assessment. Int. J. Remote Sens. 32 (15), 4407–4429. Powell, R., Roberts, D., Dennison, P., Hess, L., 2007. Sub-pixel mapping of urban land cover using multiple endmember spectral mixture analysis: Manaus, Brazil. Remote Sens. Environ. 106 (2), 253–267. Roberts, D., Gardner, M., Church, R., Ustin, S., Scheer, G., Green, R., 1998. Mapping chaparral in the Santa Monica Mountains using multiple endmember spectral mixture models. Remote Sens. Environ. 65 (3), 267–279. Roberts, D., Halligan, K., Dennison, P., 2007. VIPER Tools User Manual. UC Santa Barbara, Department of Geography, Visualization and Image Processing for Environmental Research Laboratory. Roberts, D.A., Dennison, P.E., Gardner, M.E., Hetzel, Y., Ustin, S.L., Lee, C.T., 2003. Evaluation of the potential of Hyperion for fire danger assessment by comparison to the Airborne Visible/Infrared Imaging Spectrometer. IEEE Trans. Geosci. Remote Sens. 41 (6), 1297–1310. Rosso, P., Ustin, S., Hastings, A., 2005. Mapping marshland vegetation of San Francisco Bay, California, using hyperspectral data. Int. J. Remote Sens. 26 (23), 5169–5191. Sarabandi, P., Yamazaki, F., Matsuoka, M., Kiremidjian, A., 2004. Shadow detection and radiometric restoration in satellite high resolution

71

images. Geosci. Remote Sens. In: Symposium. IGARSS'04 Proceedings, 2004 IEEE International. IEEE, pp. 3744–3747. Shawe-Taylor, J., Cristianini, N., 2000. An Introduction to Support Vector Machines: and Other Kernel-Based Learning Methods. Cambridge University Press, England. Shuster, W., Bonta, J., Thurston, H., Warnemuende, E., Smith, D., 2005. Impacts of impervious surface on watershed hydrology: a review. Urban Water J. 2 (4), 263–275. Small, C., 2003. High spatial resolution spectral mixture analysis of urban reflectance. Remote Sens. Environ. 88 (1), 170–186. Van de Voorde, T., De Roeck, T., Canters, F., 2009. A comparison of two spectral mixture modelling approaches for impervious surface mapping in urban areas. Int. J. Remote Sens. 30 (18), 4785–4806. Van der Linden, S., Hostert, P., 2009. The influence of urban structures on impervious surface maps from airborne hyperspectral data. Remote Sens. Environ. 113 (11), 2298–2305. Weng, Q., Hu, X., Liu, H., 2009. Estimating impervious surfaces using linear spectral mixture analysis with multitemporal ASTER images. Int. J. Remote Sens. 30 (18), 4807–4830. Yang, J., He, Y., Caspersen, J., 2015. Fully constrained linear spectral unmixing based global shadow compensation for high resolution satellite imagery of urban areas. Int. J. Appl. Earth Obs. Geoinf. 38, 88–98. Yang, J., He, Y., Oguchi, T., 2014. An endmember optimization approach for linear spectral unmixing of fine-scale urban imagery. Int. J. Appl. Earth Obs. Geoinf. 27, 137–146. Yuan, F., Bauer, M.E., 2006. Mapping impervious surface area using high resolution imagery: a comparison of object-based and per pixel classification. In: Proceedings of American Society for Photogrammetry and Remote Sensing Annual Conference 2006, Reno, Nevada. Zhou, W., Huang, G., Troy, A., Cadenasso, M.L., 2009. Object-based land cover classification of shaded areas in high spatial resolution imagery of urban areas: a comparison study. Remote Sens. Environ. 113 (8), 1769–1777. Zhou, W., Troy, A., 2008. An object‐oriented approach for analysing and characterizing urban landscape at the parcel level. Int. J. Remote Sens. 29 (11), 3119–3135. Zhou, Y., Wang, Y., Gold, A.J., August, P.V., 2010. Modeling watershed rainfall–runoff relations using impervious surface-area data with high spatial resolution. Hydrogeol. J. 18 (6), 1413–1423.