International Journal of Applied Earth Observation and Geoinformation
Waveform-based point cloud classification in land-cover identification
Introduction
Airborne laser scanning (ALS) data (also called topographic LiDAR data for Light Detection and Ranging data) are being used increasingly to classify land use and land cover (Wagner et al., 2008, Alexander et al., 2010, Heinzel and Koch, 2011, Crosilla et al., 2013, Dalponte et al., 2008). ALS systems apply two distinct techniques in which the return signal that is recorded is either full-waveform or discrete-return (also called multiple-return) (Alexander et al., 2010). Traditional discrete-return systems provide individual return with only discrete time measurements (Mallet et al., 2011). Full-waveform LiDAR data, which furnish more object information in the path of the laser pulse than discrete-return LiDAR data, consist of a succession of cross-section profiles of the landscape. Based on the waveform-based data, we can easily identify vertical profile of 3D object in the laser-pulse path when compared to only using echo-based (discrete-return) data.
The waveform pattern is affected by the surface reflectance, geometric structure, and roughness of the laser footprint. Amplitude, echo width, cross section, and the number of echoes (returns) are valuable parts of the information obtained from full-waveform data, which have been used successfully to classify land use and land cover (Alexander et al., 2010). For example, whereas echo amplitudes vary with the radiometric and geometric properties of targets, echo widths describe the durations of the received signal and can be applied easily in classification. Because the echoes for fields of trees are generally wider than those for roads or meadows (Wagner et al., 2008), echo width provides critical information for identifying vegetation types and detecting roads (Mallet et al., 2011, Hollaus et al., 2011). From the distance between the sensor and the target and from the amplitude and the width of an echo, the backscatter cross section can be derived, which is suitable for describing the scattering properties of targets (Alexander et al., 2010). Alexander et al. (2010) used major features such as the cross section in the classifier to distinguish between grass and roads successfully. Moreover, Crosilla et al. (2013) analyzed the skewness and kurtosis of elevation and intensity-distribution values from LiDAR data for classification.
The extracted full-waveform features can be used for classifying land cover, especially for distinguishing vegetation from other land cover types. The first echo is often acquired from the vegetation and other surface features that can be penetrated, such as the crown and trunk of trees. The last echo is a return message indicating that the last laser pulse signal is attenuating to the end, which may contain ground information or non-ground points such as dense vegetation. Jutzi and Stilla (2003) showed that vegetation and roof edges contain features that generate multiple echoes, and thus echo numbers provide key information for classifying land cover. Therefore, LiDAR waveform information on land-cover types contains single-echo (number of echoes = 1) and multi-echo (number of echoes > 1) waveform features. To consider land cover with single-echo and multi-echo waveform features along the path of a laser pulse, waveform-based classifiers were applied that included both single-echo and multi-echo features in our current study.
LiDAR features also exhibit spatial autocorrelation, which indicates that the features in a sampling point are similar to those in a neighboring sampling point. The spatial (neighborhood) information on the features of each waveform is determined using the local waveform sets included in the cylindrical neighborhood centered on that waveform (Guo et al., 2011). In the study, the average or standard deviation of the waveform characteristics was selected in the neighborhood of each waveform as the spatial features in the classifier. Furthermore, LiDAR and optical image data are generally complementary (Guo et al., 2011, Mutlu et al., 2008). Multi-source data composed of full-waveform LiDAR data and orthoimages were used for point cloud classification. The orthoimage is composed of three spectral bands in the visible domain: Red (R), Green (G), and Blue (B). The RGB channels of an orthoimage are used as three independent optical features at the nearest locations of first-echo LiDAR points in this study.
Since waveform data can be divided into two categories, one contains single major echo such as the return waveforms from open ground. The other contains multiple echoes such as the return waveforms for trees and roof edges. Waveform features were analyzed with respect to the single- and multi-echo laser-path samples to investigate whether full-waveform data can provide effective information on land-cover classifications. In this study, waveform-based classifiers were performed using support vector machine (SVM) with the waveform LiDAR features, optical image information and the spatial features of LiDAR data. Finally, the classification accuracy of the waveform-based model was compared using the conventional echo-based classifier.
Section snippets
Study area and materials
Our study area was Namasha (Namaxia), which is a famous source of precious wood in Taiwan. Namasha is a suburban district in the northeastern part of Kaohsiung City that is located upstream from the Kao-ping river watershed (Fig. 1). Namasha suffered substantial destruction during Typhoon Morakot in 2009. The entire study area was 0.95 km2, and the average elevation above the sea level and slope were approximately 722 m and 18°. Full-waveform LiDAR data were acquired from three LiDAR systems:
Methods
Fig. 2 presents a flowchart of the proposed model in this study, which comprised data processing, feature selection, and classification. First, LiDAR waveforms were decomposed into waveform components. LiDAR features can be derived from the waveform components. According to the number of echoes of a waveform, the features can be generated and categorized as single-echo features (number of echoes = 1) and multi-echo features (number of echoes >1). Major features were then selected using the
Result and discussion
Fig. 4 shows the frequency percentage of single-echo features, such as echo amplitude, width, cross section and etc. in various land cover types. Echo amplitudes of single echoes were the highest for grass and lowest for roads in the study area (Fig. 4(a)). Echo width was generally wider for grass and trees than for buildings and roads (Fig. 4(b)), which agrees with previous results (Guo et al., 2011) indicating that wide echoes correspond to vegetation because vegetation spreads the LiDAR
Conclusions
The study integrated LiDAR waveform data, orthoimage data, and the spatial features of waveform data that can be used successfully in land-cover point cloud classification. Full-waveform LiDAR data offer advantages including the ability to consider 3D structures when classifying land cover. Compared to echo-based classifiers, waveform-based classifiers were used to effectively identify various land-cover types along the path of a laser pulse. Waveform-based classifiers can improve the poor
Acknowledgement
The authors wish to thank the editors and reviewers for their valuable comments and suggestions. This research was supported by a grant from the National Science Council, R.O.C. (NSC-101-2221-E-006-181-MY3).
References (26)
- et al.
Backscatter coefficient as an attribute for the classification of full-waveform airborne laser scanning data in urban areas
ISPRS J. Photogramm. Remote Sens.
(2010) - et al.
Feature extraction based on the Bhattacharyya distance
Pattern Recognit.
(2003) - et al.
Relevance of airborne lidar and multispectral image data for urban scene classification using random forests
ISPRS J. Photogramm. Remote Sens.
(2011) - et al.
Correction of laser scanning intensity data: data and model-driven approaches
ISPRS J. Photogramm. Remote Sens.
(2007) - et al.
Exploring full-waveform LiDAR parameters for tree species classification
Int. J. Appl. Earth Observ. Geoinf.
(2011) - et al.
Synergistic use of QuickBird multispectral imagery and LIDAR data for object-based forest species classification
Remote Sens. Environ.
(2010) - et al.
Relevance assessment of full-waveform lidar data for urban area classification
ISPRS J. Photogramm. Remote Sens.
(2011) - et al.
Mapping surface fuel models using lidar and multispectral data fusion for fire behavior
Remote Sens. Environ.
(2008) - et al.
Gaussian decomposition and calibration of a novel small-footprint full-waveform digitising airborne laser scanner
ISPRS J. Photogramm. Remote Sens.
(2006) - et al.
Characterizing vertical forest structure using small-footprint airborne LiDAR
Remote Sens. Environ.
(2003)
Potential of the incidence angle effect on the radiometric calibration of full-waveform airborne laser scanning in urban areas
Am. J. Remote Sens.
Calibration of full-waveform ALS data based on robust incidence angle estimation
LiDAR data filtering and classification by skewness and kurtosis iterative analysis of multiple point cloud data categories
Appl. Geomat.
Cited by (14)
Using artificial intelligence and data fusion for environmental monitoring: A review and future perspectives
2022, Information FusionCitation Excerpt :For instance, in [396], a point-clouds-based approach is proposed to detect LULC changes from light detection and ranging (LiDAR) images, where point-clouds data has been extracted from stereo satellite imagery. Similarly, in [397], Tseng et al. introduce a Waveform-based classifier for clarifying and classifying point clouds of LULC. Moving forward, the authors in [398] propose a CNN-based LULC classification method, which includes (i) fusion of point clouds and very high-spatial resolution RS images and (ii) stratified segmentation.
Decomposition of LiDAR waveforms by B-spline-based modeling
2017, ISPRS Journal of Photogrammetry and Remote SensingCitation Excerpt :Different from the last-generation discrete-return LiDAR systems that can only record a very limited number of echoes in a laser shot, full-waveform laser scanners have the ability to record the complete waveform of the backscatter response, which offers the opportunity for researchers and end users to adopt advanced signal processing algorithms that can detect echoes with higher accuracy and reliability (Mallet and Bretar, 2009). More echo features (e.g., the pulse width and return energy) can also be extracted from waveform data (Hancock et al., 2015; Pirotti, 2011), and they have been shown to be greatly helpful in a variety of geoscience applications such as land cover classification (Fieber et al., 2013; Liu et al., 2015; Mallet et al., 2011; Tseng et al., 2015), building extraction (Michelin et al., 2012; Słota, 2015), canopy height retrieval (Gwenzi and Lefsky, 2014; Hayashi et al., 2013; Nie et al., 2015), and biomass estimation (Allouis et al., 2013; Boudreau et al., 2008; Zhuang et al., 2015). Accurately extracting echoes from LiDAR waveforms is a challenging task.
Automated cropping intensity extraction from isolines of wavelet spectra
2016, Computers and Electronics in AgricultureCitation Excerpt :In order to address these two significant challenges, this paper aims to develop an automatic Cropping Intensity extraction method based on the Isolines of Wavelet Spectra (CIIWS) obtained through continuous wavelet transform. The continuous wavelet transform has long been successfully applied for pattern recognition in agriculture and related research fields (Du et al., 2006; Gaucherel, 2002; Qiu et al., 2016, 2014a; Tseng et al., 2015; Zhang et al., 2014). These studies revealed that the wavelet spectra/features could efficiently capture the major signals of our study objects (Zhang et al., 2014).
The Use of Artificial Intelligence and Satellite Remote Sensing in Land Cover Change Detection: Review and Perspectives
2024, Sustainability (Switzerland)Airborne small-footprint full-waveform LiDAR data for urban land cover classification
2022, Frontiers in Environmental ScienceSupport Vector Machine for Land Cover Classification using Lidar Data
2021, IOP Conference Series: Earth and Environmental Science