Open Access
Add to BibSonomyAbstract
Urban land cover map is essential for urban planning, environmental studies and management. This paper aims to demonstrate the potential of geometric and radiometric features derived from LiDAR waveform and point cloud data in urban land cover mapping with both parametric and non-parametric classification algorithms. Small footprint LiDAR waveform data acquired by RIEGL LMS-Q560 in Zhangye city, China is used in this study. A LiDAR processing chain is applied to perform waveform decomposition, range determination and radiometric characterization. With the synergic utilization of geometric and radiometric features derived from LiDAR data, urban land cover classification is then conducted using the Maximum Likelihood Classification (MLC), Support Vector Machines (SVM) and random forest algorithms. The results suggest that the random forest classifier achieved the most accurate result with overall classification accuracy of 91.82% and the kappa coefficient of 0.88. The overall accuracies of MLC and SVM are 84.02, and 88.48, respectively. The study suggest that the synergic utilization of geometric and radiometric features derived from LiDAR data can be efficiently used for urban land cover mapping, the non-parametric random forest classifier is a promising approach for the various features with different physical meanings.
© 2015 Optical Society of America
1. Introduction
It is well established that urbanization process is one of the most significant impacts on land surfaces of our planet [1,2]. Urban area has become the major settlement of humans and more than 50% of the world population are now living in urban areas [3]. To provide support for decision making, accurate land cover maps are needed in urban planning. Since remote sensing technology is a rapid and efficient solution for acquiring image data set over land surfaces, it has been widely used in urban planning and environmental management [4,5]. Remotely sensed data not only provides base map for urban planning, it also can be applied to produce thematic maps over urban area. Usually, urban planning requires fine resolution land cover products, and high-resolution optical remote sensing images have been adopted for generating urban land cover maps. To date, the large volume passive optical images acquired by high resolution sensors such as QuickBird and IKONOS have been widely used in urban land cover mapping [6–8]. However, the displacement and shadows in the high resolution optical images make it difficult to generate accurate land cover maps in urban areas.
LiDAR (LIght Detection And Ranging) is an active optical sensor which integrates laser ranging, Global Positioning System (GPS) and Inertial Navigation System (INS) to capture both geometric and radiometric information of targets. It provides direct range measurements of land surfaces by receiving the signal reflected by illuminated targets. In the past decades, LiDAR has been widely applied in terrain modeling, estimation of forest vertical parameters as well as land cover mapping [9,10]. Hodgson et al. [11] introduced the synergistic use of LiDAR and color aerial imagery for urban parcel imperviousness mapping. Three algorithms 1). maximum-likelihood; 2). ISODATA spectral clustering, and 3). rule-based classification with the machine learning tool were tested to assess the LiDAR data based urban land cover mapping. The results of the study indicated that the combination aerial imagery and LiDAR data could improve the accuracy of imperviousness estimation, the rule-based classification produced the highest accuracy among the tested approaches. Brennan et al. [12] applied height and intensity of LiDAR data to classify land cover using an object-oriented approach, features derived from the LiDAR points included DSM, DEM, intensity, multiple echoes, and normalized height. The derived features were segmented and classified by an object rule based classification. Antonarakis et al. [13] introduced an object-based methods of forest and ground types classification with airborne LiDAR. In the study, elevation and intensity were integrated for object segmentation and classification, the result indicated that object based approach is an efficient way to generate land cover map using LiDAR data. Mallet et al. [9] conducted a comprehensive assessment of features derived from LiDAR data for urban land cover classification, features derived from waveform and features those considered spatial relationship of point cloud were examined for urban land cover classification. Three approaches including F-score, ReliefF, and SVM–RFE were adopted to assess the performance of the features in discriminating urban land cover types. The MLC algorithm was examined for LiDAR land cover mapping studies by Yan et al. [14] and Yan & Shaker [15], the radiometrically corrected LiDAR data was applied for land cover mapping. However, there is a lack of comparison of different classifiers on LiDAR data classification.
In general, LiDAR data provides detailed elevation information of land surfaces, the elevation change pattern of targets has lot of information which can be used for land cover classification. In addition, the radiometric attributes also can be adopted for land cover classification. LiDAR data is commonly delivered in point cloud, while the associated intensity values of LiDAR points denote the amplitude of return energy. Although many efforts have been made to correct LiDAR intensity for radiometric characterization of targets [16,17], it is difficult to calibrate intensity to an uniform attribute like reflectance since the different sensor configurations and different atmospheric situation during the data acquisition. Aims to propose a comprehensive robust processing chain for LiDAR waveform data, we developed series algorithms for LiDAR waveform processing, including waveform decomposition, range determination and radiometric characterization.
The objective of this paper is to demonstrate the application of urban land cover classification using geometric and radiometric features derived from both LiDAR waveform and point cloud data. With the utilization of a parametric classifier, i.e. MLC, and two nonparametric classifiers, i.e, SVM and random forest, this paper is expected to address the basic issue in LiDAR data classification with different unit/unitless features with various data range. Based on the algorithms developed in our previous studies, LiDAR waveform data was decomposed by Gaussian Mixture Model (GMM), location of each waveform was determined by the range determination algorithm. Features derived from both LiDAR waveform and point cloud including geometric and radiometric attributes were adopted to discriminate urban land cover types. The MLC, SVM and random forest algorithm were selected to implement the supervised land cover classification. The comparison analysis was also carried out to assess the performance of different classifiers.
2. Study site and data set
LiDAR data used in this paper was acquired by LMS Q560 system during WATER experiment [18], the study area was located in Zhangye city, Gansu province (Fig. 1). Since the objective of this paper is to investigate the potential of utilizing LiDAR features for deriving urban land cover map, LiDAR waveform data over Zhangye downtown was selected to carry out the experiment.
There were thirteen flight strips covering the downtown area of Zhangye city, the average flight height of the mission was about 700 meters above ground with the average density of waveform data was ~1.7bins/m2, These LiDAR waveform data were imported from the raw data and calibrated by GeocodeWF, which is a standalone software for handling raw waveform data collected by RIEGL LMS-Q560 laser scanner-based LiDAR systems into geocoded points in a projected coordinate system [19]. In the WATER experiment, LiDAR waveform data was in GeocodeWF binary geocoding and calibrated waveform file format.
3. Methods
Figure 2 illustrates the processing workflow, consisting of waveform decomposition, range determination, radiometric characterization, feature extraction, land cover classification as well as accuracy assessment, the details of the steps are described in the following sub-sections.
3.1 Waveform decomposition and range determination
Gaussian decomposition has been widely applied in LiDAR waveform processing, the method is based on the assumption that the energy of transmitted LiDAR waveform yields to a Gaussian distribution, the returned LiDAR waveform is the convolution between transmitted LIDAR waveform and response of illuminated target [20]. The returned LiDAR waveform can be approximated by a GMM according to the attribute of signal convolution, it consequently be described as:
where is the returned LiDAR waveform with k Gaussian components. is a single Gaussian component with parameters ,,, these parameters denote the amplitude, position and width of a single Gaussian component.To model LiDAR waveform with “good” Gaussian model, a curve fitting based approach has been developed to identify the number of Gaussian components and also provide initial parameters for Gaussian decomposition [20]. As a promising solution for LiDAR waveform decomposition, the approach was applied for Gaussian decomposition in this study.
It is well established that the complex terrain change in LiDAR footprint tends to induce the pulse deformation in LiDAR waveform [21]. Based on the simulation of pulse deformation over complex terrain area, an approach for range determination correction was developed for generating point cloud from small footprint LiDAR waveform data [22]. The approach based on the Gaussian decomposition of LiDAR waveform data, it has been proved that the resulted point cloud data would be more accurate in characterizing the height of Illuminated target, especially for tree canopies. In this study, the approach was adopted to estimate the geolocation of LiDAR waveforms
3.2 Feature extraction
Although many features have been proposed by existing studies, there is no widely accepted standard route for LiDAR waveform based urban land cover classification. This study starts from LiDAR waveform data, with the consideration of existing studies and also the demonstration of our processing chain in LiDAR applications, only a few features were selected for discriminating urban land cover types in this study:
- ●ΔZ: height difference between the current point and the lowest point found in cylindrical volume with a fixed radius, in this study the radius was empirically set to 20m [9].
- ●σZ: it represents the variation of height values which falls into a vertical circle, in this study the radius of the circle was 10m [9].
- ●Gaussian parameters of decomposed LiDAR waveform: hi, wi are the amplitude and width of peak in waveform acquired by the Gaussian decomposition [20].
- ●NRF: Reflectance-like feature derived from LiDAR waveform data, with the correction of the multi-peak waveform radiometric correction, NRF is normalized by the emitted waveform as well as the atmospheric attenuation. The detailed description can be found in [23].
The coordinates of LiDAR point cloud are generated from the Gaussian parameters of waveform decomposition. Each point obtained from LiDAR waveform has five features for discriminating different land cover types. For a given point derived from a particular LiDAR waveform, the land cover information could be derived from the corresponding features. Generally, land cover map is delivered in either raster or vector-polygon format. In this study, the natural neighbor interpolation method was adopted to generate the raster format features of LiDAR data. The density of LiDAR point cloud used in this study was 2.14 points/m2, the pixel size of the raster image was set as 1.0 m.
3.3 Classification methods
Classification algorithms are vital research topic in machine learning community. For remote sensing data classification, many studies have been conducted to examine different classifiers for LiDAR data based land cover mapping [9,24,25]. It is well addressed that LiDAR data could be applied for land cover mapping. In this study, features derived from LiDAR data are with different range without comparable units. To investigate the performance of different classifier in discriminating urban land cover types by using the features derived by our approach, Maximum Likelihood Classification (MLC), Support Vector Machine (SVM) and random forest algorithm were adopted to perform urban land cover classification.
MLC is a conventional method for remote sensing data classification. The basic assumption of MLC is that the data is normally (Gaussian) distributed, with the estimation of Gaussian model for the training data set, the discriminant function is identified for sample classification. It has been well accepted that the MLC is a promising classification approach for normal distribution data set the 1980s [26,27], especially for spectral image classification.
Decision tree algorithm was introduced to create a model for predicting the value of target attributes with input variables. Over the past decades, many decision tree learning methods were proposed to automatically construct a decision tree from training data set, e.g. Quinlan [28] developed Iterative Dichotomiser 3 (ID 3) method with the extended version C4.5 to generate a decision tree from a training samples using the principle of the maximum entropy [29], these methods were developed to construct a single decision tree for predict values of dependent variables. Other type of decision tree algorithms are designed by using multiple decision trees, these algorithms are named as ensemble decision tree algorithms. Random forest is an ensemble decision tree algorithm proposed by Breiman [30]. As a non-parametric approach for classification and prediction, random forest is implemented by constructing multiple many simple trees at training time, the prediction is made by considering the output of each individual tree. The advantages of random forest method are: 1) producing an unbiased estimation of the error as the forest construction progresses; 2) handles variables with different types (e.g. categorical, numerical) in same data set for constructing decision tree and 3) the decision tree learning algorithm constructs the decision rules by combining many single small trees, which would reduces the possibility of data set over-fitting. Consequently, random forest has attracted a rising interests over the past decades. In terms remote sensing data processing, many studies indicated that the random forest is an efficient approach for feature selection, image classification and estimation of biophysical parameters [9, 31].
Support Vector Machines (SVM) is a supervised machine learning algorithm for data classification and regression analysis [32]. In the past decade, SVM algorithms have been widely applied in remote sensing image classification [33–36]. As a non-parametric algorithm, it performs classification for the data sets which could not be modeled by standard parametric probability density functions. For a non-linearly separable data set, SVM algorithm projects the original data set to a higher dimensional space. With the transformation of feature space, the original data set would become separable in the high dimensional feature space. Aims to find a map function to project the original data to high dimensional feature space, a kernel function is adopted to find the hyper-plane to maximize the distance between the closest training samples for different classes. More technical about SVM algorithm details can be found from [37].
Generally, the major land cover types over urban area are building, road, bare ground, tree, water and grass. Most existing studies consider building, bare ground, tree and water body since these are dominant land cover types in most cities and can be discriminated by LiDAR data. Considering the situation of land cover types in the study area, only building, ground, and tree are selected as three dominant land cover types in the study area.
3.4 Sample collection and accuracy assessment
In this study, three supervised classifiers were adopted for discriminating urban land cover types, image truth was collected for both classifier training and the accuracy assessment of results. Based on the visual interpretation upon the images acquired by CCD camera and the five feature layers, Regions Of Interest (ROIs) of ground, tree and building were collected by manual digitalization over the central area of the downtown in Zhangye city (Table 1) . To ensure the ROIs are randomly distributed over the study area, the training samples and the validation samples were selected separately.
Table 1. Number of ROIs used in the classification
For remote sensing data based thematic mapping, the accuracy report is used typically to describe the degree of “correctness” of a map or classification [38]. It is well accepted that the analysis of confusion matrix is a promising way to carry out the accuracy assessment of remote sensing classification [38–40]. The land cover classification in this study was carried out at pixel level, and validation ROIs were used to construct the confusion matrix. Indices which describe commission and omission errors were calculated to provide the accuracy report of land cover classification with the confusion matrix.
4. Results
Aerial imagery was simultaneously acquired with LiDAR data in the WATER experiment, to carry out the visual comparison between the land cover classification results, the aerial imagery was used to illustrate the detailed scene of the study area. Since each aerial image had a limited spatial coverage, with the consideration of keeping balance between the figure quality and spatial coverage, the coverage of an aerial image over the central area of downtown was identified for visualizing the results Fig. 3 shows the selected aerial image over the center of downtown, the image was acquired in June, 2008. As described in previous sections, five feature layers including ΔZ, σZ, hi, wi and NRF were derived from LiDAR waveform and point cloud. Figure 4, Fig. 5, Fig. 6, Fig. 7 and Fig. 8 shows the corresponding feature layers to highlight the relationship between the features of different land cover types.
The visual comparison among the feature layers and the RGB image suggests that the selected features derived from LiDAR highlight different land cover types. Specially, it is observed that the ΔZ highlights building roof, while the wi highlights road trees which provides additional confirmation for the range correction method, and the NRF combines the characterizations of hi, wi. In terms of σZ, the borders of building and tree are enhanced since the height variation of these areas are relatively high.
Based on the above feature layers, supervised classification was carried out to derive urban land cover maps. As mentioned in the previous section, MLC, SVM and random forest algorithms were applied to implement the land cover discrimination. Figure 9, Fig. 10, and Fig. 11 are the land cover maps generated by MLC, SVM and random forest algorithm over the area illustrated in Fig. 3. Visual interpretation upon the land cover maps suggests that the outline of study area is characterized by the there classifiers. The comparison among the maps indicates the MLC method classified a lot of building pixels as ground, while the SVM and random forest method generated similar land cover maps.
Fig. 9 Land cover map obtained by MLC classifier for the corresponding area in Fig. 3 (1.0m spatial resolution).
Fig. 10 Land cover map obtained by SVM classifier for the corresponding area in Fig. 3(1.0m spatial resolution).
Fig. 11 Land cover map obtained by random forest algorithm for the corresponding area in Fig. 3(1.0m spatial resolution).
The quantitative accuracy assessment of land cover maps was also conducted to provide quality information of the land cover maps. Table 2 summarizes the overall accuracies and Kappa coefficients for the results of MLC, SVM and RF classification. It is observed that the RF achieved the highest overall accuracy and Kappa coefficient among the three classifiers. The overall accuracy and Kappa coefficient of the MLC classification are 84.02%, 0.76 respectively.
Table 2. Summary of accuracy assessment
The confusion matrix of the land cover maps generated by the MLC, SVM and random forest method is presented in Table 3. For the three land cover classification results, it is observed that the class of ground has relatively low user accuracy, the producer accuracies of building are lower than both ground and tree, which means the building and ground confused with each other. Specially, the visual inspection upon the land cover classification suggests that a lot of actual buildings are classified as ground in MLC. Based on the visual inspection and quantitative assessment, we can conclude that the random forest achieved the most accurate classification results among the three classifiers.
Table 3. Confusion matrix for the MLC, SVM and random forest classifications
5. Discussion and conclusion
This paper presents an application of LiDAR data for urban land cover mapping, the motivation of this study is expected to demonstrate the application of urban land cover classification using LiDAR data and compare the performance of different classification algorithms for discriminating land cover types. The workflow involves a processing chain for LiDAR data processing, it consists of waveform decomposition, range determination and radiometric characterization. Features derived from both waveform and point cloud were adopted for land cover classification. The performance of MLC, SVM and random forest algorithms in generating LiDAR-based urban land cover maps was examined. The experiments were carried out with LiDAR data acquired in Zhangye city, China. Visual inspection upon the classification results showed that the proposed processing chain worked well. The assessed accuracy inferred that the random forest achieved the most accurate results among the three classifiers, with overall accuracy of 91.82% and Kappa of 0.88.
Generally, the parametric classification algorithms apply distance functions which combine features of data set for sample discrimination. The features derived from LiDAR data have various data ranges with different units or unitless, for such data set, although normalization strategies can be applied in the process of distance calculation, it is difficult to model the distribution of features with parametric models. Consequently, the parametric classification algorithm(i.e. MLC) may not work well in discriminating samples with hybrid features. In contrast, the non-parametric classifiers (i.e. SVM and random forest) implement sample classification without assumption to prior distribution, data set with hybrid features can be classified without fitting features using parametric models. The results of this study suggest that the non-parametric classification algorithms SVM and random forest achieved higher accuracies than the MLC approach and the random forest algorithm is a promising approach for the classification using the unit/unitless features with different data range.
The existing studies have examined number of features derived from LiDAR data for land cover classification. Since one of the important objectives was to address the selection of classifier to implement the hybrid-feature classification, this study only selected five features for discriminating different land cover types. However, more features would be assessed to identify the optimal combination of features for characterizing land surfaces. Although the radiometric attributes of LiDAR can be used for land cover classification, these features only provide the radiometric property for single band, and usually do not provide adequate separability for specific land cover types, e.g. grassland and bare ground have very similar geometric attributes with significant different spectral, but it is difficult to discriminate these two types only using LiDAR data. Consequently, with the fusion of spectral attributes obtained from passive optical images, these land cover types could be discriminated with high accuracy.
Further studies are expected to implement urban land cover mapping using the NRF with additional features derived from LiDAR data and investigate the capability of different classifiers in discriminating land cover types, and therefore generate a high accuracy 3D urban land cover map. The final map would be a 3D classification map that provides an explicit 3D view for urban planning and management.
Acknowledgments
This work was supported by Hundred Talents Program of Chinese Academy of Sciences, the National Natural Science Foundation of China (41471294) and the Major State Basic Research Development Program of China(2013CB733405). The authors would like to thank the Watershed Allied Telemetry Experimental Research for providing LiDAR data.
References and links
1. P. M. Vitousek, H. A. Mooney, J. Lubchenco, and J. M. Melillo, “Human domination of Earth's ecosystems,” Science 277(5325), 494–499 (1997). [CrossRef]
2. L. Zhou, R. E. Dickinson, Y. Tian, J. Fang, Q. Li, R. K. Kaufmann, C. J. Tucker, and R. B. Myneni, “Evidence for a significant urbanization effect on climate in China,” Proc. Natl. Acad. Sci. U.S.A. 101(26), 9540–9544 (2004). [CrossRef] [PubMed]
3. Population Division, United Nations Department of Economic and Social Affairs, (UN DESA), 2010, World Urbanization Prospects: The 2009 Revision. Highlights, New York, NY, USA.
4. M. K. Ridd, “Exploring a VIS (vegetation-impervious surface-soil) model for urban ecosystem analysis through remote sensing: comparative anatomy for cities,” Int. J. Remote Sens. 16(12), 2165–2185 (1995). [CrossRef]
5. J. Xiao, Y. Shen, J. Ge, R. Tateishi, C. Tang, Y. Liang, and Z. Huang, “Evaluating urban expansion and land use change in Shijiazhuang, China, by using GIS and remote sensing,” Landsc. Urban Plan. 75(1), 69–80 (2006).
6. Y. Ban, H. Hu, and I. M. Rangel, “Fusion of Quickbird MS and RADARSAT SAR data for urban land-cover mapping: object-based and knowledge-based approach,” Int. J. Remote Sens. 31(6), 1391–1410 (2010). [CrossRef]
7. A. P. Carleer and W. Eeleonore, “Urban land cover multi‐level region‐based classification of VHR data by selecting relevant features,” Int. J. Remote Sens. 27(6), 1035–1051 (2006). [CrossRef]
8. W. Zhou, G. Huang, A. Troy, and M. L. Cadenasso, “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 (2009). [CrossRef]
9. C. Mallet, F. Bretar, M. Roux, U. Soergel, and C. Heipke, “Relevance assessment of full-waveform lidar data for urban area classification,” ISPRS J Photogramm. 66(6), 71–84 (2011). [CrossRef]
10. W. Y. Yan, A. Shaker, and N. El-Ashmawy, “Urban land cover classification using airborne LiDAR data: a review,” Remote Sens. Environ. 158, 295–310 (2015). [CrossRef]
11. E. Hodgson, J. R. Jensen, J. A. Tullis, K. D. Riordan, and C. M. Archer, “Synergistic use of lidar and color aerial photography for mapping urban parcel imperviousness,” Photogramm Eng Rem S 69(9), 973–980 (2003). [CrossRef]
12. R. Brennan and T. L. Webster, “Object-oriented land cover classification of lidar-derived surfaces,” Can. J. Rem. Sens. 32(2), 162–172 (2006). [CrossRef]
13. S. Antonarakis, R. Keith, and J. Brasington, “Object-based land cover classification using airborne LiDAR,” Remote Sens. Environ. 112(6), 2988–2998 (2008). [CrossRef]
14. W. Y. Yan and A. Shaker, “Radiometric correction and normalization of airborne LiDAR intensity data for improving land cover classification,” IEEE T Geosci Remote. 52(12), 7658–7673 (2014). [CrossRef]
15. W. Y. Yan, A. Shaker, A. Habib, and A. P. Kersting, “Improving classification accuracy of airborne LiDAR intensity data by geometric calibration and radiometric correction,” ISPRS J Photogramm. 67, 35–44 (2012). [CrossRef]
16. B. Höfle and N. Pfeifer, “Correction of laser scanning intensity data: Data and model-driven approaches,” ISPRS J Photogramm. 62(6), 415–433 (2007). [CrossRef]
17. S. Kaasalainen, H. Hyyppa, A. Kukko, P. Litkey, E. Ahokas, J. Hyyppa, H. Lehner, A. Jaakkola, J. Suomalainen, A. Akujarvi, M. Kaasalainen, and U. Pyysalo, “Radiometric calibration of LIDAR intensity with commercially available reference targets,” IEEE T Geosci Remote. 47(2), 588–598 (2009). [CrossRef]
18. X. Li, X. Li, Z. Li, M. Ma, J. Wang, Q. Xiao, and Q. Liu, “Watershed allied telemetry experimental research,” J Geophys Res-Atmos 114(D22) (2009). [CrossRef]
19. WWW1, http://geolas.com
20. Y. Qin, T. T. Vu, and Y. Ban, “Toward an Optimal Algorithm for LiDAR Waveform Decomposition,” IEEE Geosci Remote 482–486(3), 9 (2012).
21. B. Jutzi and U. Stilla, “Range determination with waveform recording laser systems using a Wiener Filter,” ISPRS J Photogramm. 61(2), 95–107 (2006). [CrossRef]
22. Y. Qin, T. T. Vu, Y. Ban, and Z. Niu, “Range determination for generating point clouds from airborne small footprint LiDAR waveforms,” Opt. Express 20(23), 25935–25947 (2012). [CrossRef] [PubMed]
23. Y. Qin, W. Yao, T. T. Vu, S. Li, Z. Niu, and Y. Ban, “Characterizing Radiometric Attributes of Point Cloud Using a Normalized Reflective Factor derived from Small Footprint LiDAR WaveformRange determination for generating point clouds from airborne small footprint LiDAR waveforms,” IEEE J Sel Top Appl. 8(2), 740–749 (2015).
24. B. Höfle, M. Hollaus, and J. Hagenauer, “Urban vegetation detection using radiometrically calibrated small-footprint full-waveform airborne LiDAR data,” ISPRS J Photogramm. 67, 134–147 (2012). [CrossRef]
25. T. Sasaki, I. Junichi, I. Keiko, M. Yukihiro, and K. Katsunori, “Object-based classification of land cover and tree species by integrating airborne LiDAR and high spatial resolution imagery data,” Landsc Ecol Eng. 8(2), 157–171 (2012). [CrossRef]
26. A. Strahler, “The use of prior probabilities in maximum likelihood classification of remotely sensed data,” Remote Sens. Environ. 10(2), 135–163 (1980). [CrossRef]
27. G. Foody, N. Campbell, M. Trodd, and F. Wood, “Derivation and applications of probabilistic measures of class membership from the maximum-likelihood classification,” Photogramm Eng Rem S 58(9), 1335–1341 (1992).
28. R. Quinlan, “Induction of decision trees,” Mach. Learn. 1(1), 81–106 (1986). [CrossRef]
29. R. Quinlan, C4. 5: programs for machine learning (Morgan kaufmann, 1993).
30. L. Breiman, “Random forests,” Mach. Learn. 45(1), 5–32 (2001). [CrossRef]
31. A. Bosch, Z. Andrew, and M. Xavier, “Image classification using random forests and ferns,” IEEE 11th International Conference on Computer Vision. 1–8 (2007). [CrossRef]
32. C. Cortes and V. Vladimir, “Support-vector networks,” Mach. Learn. 20(3), 273–297 (1995). [CrossRef]
33. M. Brown, L. Hugh, and G. Steve, “Linear spectral mixture models and support vector machines for remote sensing,” IEEE T Geosci Remote. 38(5), 2346–2360 (2000). [CrossRef]
34. G. Zhu and B. Dan, “Classification using ASTER data and SVM algorithms; The case study of Beer Sheva, Israel,” Remote Sens. Environ. 80(2), 233–240 (2002). [CrossRef]
35. F. Melgani and B. Lorenzo, “Classification of hyperspectral remote sensing images with support vector machines,” IEEE T Geosci Remote. 42(8), 1778–1790 (2004). [CrossRef]
36. D. Tuia, V. Michele, C. Loris, K. Mikhail, and M. Jordi, “A survey of active learning algorithms for supervised remote sensing image classification,” IEEE J Sel Top Appl. 5(3), 606–617 (2011). [CrossRef]
37. N. Cristianini and J. Shawe-taylor, An Introduction to Support Vector Machines and Other Kernel-based Learning Methods (Cambridge University Press, 2000)
38. G. Foody, “Status of land cover classification accuracy assessment,” Remote Sens. Environ. 80(1), 185–201 (2002). [CrossRef]
39. L. Janssen and J. W. Frans, “Accuracy assessment of satellite derived land-cover data: a review,” Photogramm. Eng. Remote Sensing 60(4), 419–426 (1994).
40. M. Herold, P. Mayaux, C. Woodcock, A. Baccini, and C. Schmullius, “Some challenges in global land cover mapping: An assessment of agreement and accuracy in existing 1 km datasets,” Remote Sens. Environ. 112(5), 2538–2556 (2008). [CrossRef]
41. M. Ali, “Roles and challenges of urban design,” J. Urban Des. 11(2), 173–193 (2006). [CrossRef]
42. S. George and G. Vosselman, “Experimental comparison of filter algorithms for bare-Earth extraction from airborne laser scanning point clouds,” ISPRS J Photogramm. 59(1), 85–101 (2004).
43. P. Justin and R. Schowengerdt, “A detailed comparison of backpropagation neural network and maximum-likelihood classifiers for urban land use classification,” IEEE T Geosci Remote. 33(4), 981–996 (1995). [CrossRef]
44. WWW2, http://geolas.com/Downloads/uewAdMxq/WaveformFileFormat.pdf
