928
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, VOL. 10, NO. 4, JULY 2013
An Object-Based Approach for Urban Land Cover Classification: Integrating LiDAR Height and Intensity Data Weiqi Zhou
Abstract—Digital surface models (DSMs) derived from light detection and ranging (LiDAR) data have been increasingly integrated with high-resolution multispectral satellite/aerial imagery for urban land cover classification. Fewer studies, however, have investigated the usefulness of LiDAR intensity in aid of urban land cover classification, particularly in highly developed urban settings. In this letter, we use an object-based classification approach to investigate whether a combination of LiDAR height and intensity data can accurately map urban land cover. We further compare the approach to a method that uses multispectral imagery as the primary data source, but LiDAR DSM as ancillary data to aid in classification. The study site is a suburban area in Baltimore County, MD. The LiDAR data were acquired in March 2005, from which DSM and two intensity layers (first and last returns), with 1-m spatial resolution were generated, respectively. Four classes were included: 1) buildings; 2) pavement; 3) trees and shrubs; and 4) grass. Our results indicated that the objectbased approach provided flexible and effective means to integrate LiDAR height and intensity data for urban land cover classification. A combination of the LiDAR height and intensity data proved to be effective for urban land cover classification. The overall accuracy of the classification was 90.7%, and the overall Kappa statistics equaled 0.872, with the user’s and producer’s accuracies ranging from 86.8% to 93.6%. The accuracy of the results were far better than those using multispectral imagery alone, and comparable to using DSM data in combination with high-resolution multispectral satellite/aerial imagery. Index Terms—Baltimore, high-resolution imagery, intensity, light detection and ranging (LiDAR), normalized digital surface model (nDSM), object-based image analysis, urban land cover classification.
I. Introduction CCURATE and timely information about urban land cover is essential for urban land management, planning, and landscape pattern analysis. Remote sensing provides the primary source of data for urban land cover mapping. As an
A
Manuscript received November 19, 2012; revised January 8, 2013; accepted February 27, 2013. Date of publication April 12, 2013; date of current version May 27, 2013. This work was supported in part by the Chinese Academy of Sciences One Hundred Talented Program, the State Key Laboratory of Urban and Regional Ecology, the Ministry of Environmental Protection of China under Grant STSN-12-01, and the National Science Foundation LTER Program (DEB 042376). The author is with the State Key Laboratory of Urban and Regional Ecology, Research Center for Eco-Environmental Sciences, Chinese Academy of Sciences, Beijing 100085, China (e-mail:
[email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/LGRS.2013.2251453
urban environment is extremely complex and heterogeneous, very high spatial resolution remotely sensed data are needed to adequately characterize the fine-scale spatial heterogeneity of urban landscape [1]. Consequently, high spatial resolution satellite and aerial imagery has been frequently used for detailed urban land cover mapping [2]–[4]. The recent availability of airborne light detection and ranging (LiDAR) data provides new opportunities for detailed urban land cover mapping at very fine scales. LiDAR is an active remote sensing technology, operating in the visible or near-infrared region of the electromagnetic spectrum. With the recent advances of airborne LiDAR technology, there is increasing interest in applying LiDAR data to urban land cover classification. LiDAR point clouds can be directly used for urban feature extraction [5]. More frequently, however, LiDAR points are first interpolated into raster layer(s), and then combined with high-resolution satellite/aerial imagery for detail urban land cover mapping. Researchers commonly used surface height information, or digital surface model (DSM) derived from LiDAR data as ancillary data to aid in classification [1], or as the primary data for classification [4], [6]. Studies have shown that the accuracy of urban land cover classification can be greatly improved by integrating multispectral imagery with LiDAR data [1], [4], [6]. In addition to height data, LiDAR also provides intensity data that reflect the material characteristics of land cover features, which can be potentially used for urban land cover classification [6], [7]. While LiDAR intensity data have been increasingly used in forest-type classification [8], only a few very recent studies have used LiDAR intensity as ancillary data to aid in urban land cover mapping [4], [6]. Few studies have investigated the usefulness of LiDAR data alone, i.e., a combination of LiDAR height and intensity information, in urban land cover classification [7], [9], particularly in highly developed urban settings, where classification is more challenging due to the fine-scale complexity of urban land cover features. This letter aims to fill this gap. Paralleled with the increasing availability of LiDAR data are the advances in object-based image analysis (OBIA), an image classification approach that has gained wide acceptance in fine-scale urban land cover mapping [10]. Rather than classifying individual pixels, object-based classification segments the imagery into objects. Consequently, in addition to spectral response, object characteristics, such as shape and spatial
c 2013 IEEE 1545-598X/$31.00
ZHOU: OBJECT-BASED APPROACH FOR URBAN LAND COVER CLASSIFICATION
Fig. 1.
929
Study site: a suburban area in the Baltimore County, MD, USA.
relations, can be used for classification [1], [10]. Many studies have shown that OBIA techniques are superior to pixel-based approaches for land cover classification from high-resolution imagery [10]. In this letter, we used an object-based classification approach to investigate whether a combination of the LiDAR height and intensity data can accurately characterize and map urban land cover. We further compared this approach to a method that used imagery as the primary data source, but LiDAR height data as ancillary data for classification.
II. Data and Methodology A. Study Site The study site was a suburban area in Baltimore County, MD, USA (Fig. 1). Land use of this area was dominated by medium- to high-density residential development, mixed with small proportions of commercial and other institutional land uses. Land cover features are typical of those in urban and suburban environments, including detached and multifamily houses, commercial buildings, paved surfaces, and vegetation cover. Therefore, the variety of the land use/land cover in the study site makes it well suited for the goal of this letter. B. Data Preprocessing 1) LiDAR Data: The LiDAR data were acquired in March 2005. Both the first and last vertical returns were recorded for each laser pulse, with the average point spacing of approximately 1 m. The returns from bare ground and nonground (e.g., tree canopy, building roofs) were separated. The LiDAR point clouds were processed to generate three separate raster datasets: a normalized digital surface model (nDSM), and two intensity image layers. Normalized digital surface model: The points returns from bare ground were interpolated into 1-m spatial resolution digital elevation model (DEM), and all returns (i.e., both returns from bare ground and nonground) into 1-m resolution DSM, using the natural neighbor interpolation method available in
Fig. 2. Subset of the aerial imagery, LiDAR data layers, and classification results. (a) Multispectral emerge imagery. (b) nDSM. (c) First return intensity. (d) Last return intensity. (e) Classification results from Method 1 (shades of gray from light to dark: grass, trees and shrubs, pavement, and buildings). (f) Classification results from Method 2.
ArcGIS 3-D analyst. The surface cover height model (referred to as nDSM) was then generated by subtracting the DEM from the DSM (Fig. 2). Intensity layers: Two intensity layers were generated from the first and last return measurements, respectively (Fig. 2). The natural neighbor interpolation method in ArcGIS 3-D analyst was used to generate the two 1-m spatial resolution intensity layers. The mean and standard deviation of intensity from first returns were 7.83 and 5.99, respectively, with the range of 0.10 to 472.82; and those from last returns were 9.17 and 5.31, respectively, with the range of 0.14 to 502.32. 2) High-Resolution Imagery Data: Color-infrared digital aerial image data with a pixel size of 0.6 m acquired in 2004 were used in this letter for comparison purposes (Fig. 2). The imagery was 3-band color-infrared, with green (510–600 nm), red (600–700 nm), and near-infrared bands (800–900 nm). The imagery data has an 8-bit radiometric depth, and was orthorectified [2]. C. Land Cover Classification In this letter, four land cover classes were identified: 1) buildings; 2) pavement; 3) trees and shrubs; and 4) grass, which are the most typical land cover types in urban and suburban landscapes. Two methods were applied to perform the land cover classification. Both methods used an OBIA
930
approach for classifications, which was implemented using the software eCognition. The classification procedures, as detailed below, however, were slightly different because the primary data sources for classification within the two methods were different. 1) Method 1: Classification Using LiDAR Data Alone: Method 1 used LiDAR data alone for classification. A rule set, a sequence of processing commands/algorithms, was developed to generate image segments and to classify them into desired land cover classes [1], [2]. Specifically, a contrast-split segmentation algorithm was first used to separate tall objects from short objects based on nDSM [6]. The minimum and maximum thresholds were set as 6 feet and 9 feet, respectively. We then further separated tall objects into buildings and trees, and classified short objects into pavement and grass, as detailed follows. Classification of tall objects into buildings and trees. A multiresolution segmentation was run for the tall objects, using both nDSM and the intensity layer from the first return measurements. The multiresolution segmentation algorithm initialized with each pixel previously classified as tall objects in the image as a separate segment, which was merged with neighboring segments based on their level of similarity in subsequent steps. The process stops when there are no more possible merges given a defined scale parameter. The scale parameter specifies the maximum heterogeneity that is allowed within each object, which indirectly controls the size of objects. The greater the scale parameter, the larger the average size of the objects. The user can also specify color and shape parameters to change the relative weighting of reflectance and shape in defining segments. In this letter, the scale parameter was set as 10 to conduct the segmentation at a very fine scale. The color criterion was given a weight of 0.9, while the shape was assigned with the remaining weight of 0.1, giving equal weights to compactness (i.e., 0.05) and smoothness. The scale parameter of 10 and the values for the color and shape parameters were determined by visual inspection of the image segmentation results, where objects were considered to be internally homogenous, i.e., all pixels within an image object belonged to one cover class [1]. Following the segmentation, tall objects were classified as buildings if the difference in intensity from the first and last returns was less than 1, and standard deviation of nDSM was less than 6 feet. In addition, tall objects that share boundaries with these previously classified buildings were further identified as buildings, if: 1) the standard deviation of nDSM
