Urban vegetations have infinite proven benefits for urban inhabitants including providing shade, improving air quality, and enhancing the look and feel of communities. But creating a complete inventory is a time consuming and resource intensive process. The extraction of urban vegetation is a challenging task, especially to monitor the urban tree heights. In this study we present an efficient extraction method for mapping and monitoring urban tree heights using fused hyperspectral image and LiDAR data. Endmember distribution mapping using the spectral angle mapper technique is employed in this study. High convenience results achieved using fused hyperspectral and LiDAR data from this semiautomatics technique. This method could enable urban community organizations or local governments to map and monitor urban’s tree height and its spatial distribution.
Remotely-sensed data and imagery provide a comprehensive, scalable means for detecting and quantifying land use land cover (LULC) change, and its use in mapping urban growth, estimating population density, and modelling sustainability and quality of life is becoming increasingly popular as the scale, cost, and spatial-temporal coverage improves [
The Light Detection and Ranging (LiDAR) and hyperspectral data provide the best opportunity in order to distinguish urban features from urban vegetations (e.g., impervious surface, and vegetated area). LiDAR airborne laser scanner has emerged as an increasingly popular tool for collecting very high-resolution structural data representing the vertical dimension of the Earth’s surface by measuring the travel time of laser pulses between the sensor and earth objects [
Reitberg [
This study examines the advantage of LiDAR and hyperspectral data fusion to improve mapping and monitoring of urban tree heights along different urban features over the University of Houston and the neighbouring urban area, Houston, USA.
A hyperspectral image and a LiDAR derived Digital Surface Model (DSM), both at the same spatial resolution (2.5 m). The hyperspectral imagery consists of 144 spectral bands in the 380 nm to 1050 nm region and has been calibrated to at-sensor Spectral Radiance Units (SRU) = µW/(cm2 sr nm). The corresponding co-Registered DSM consists of elevation in meters above sea level (per the Geoid 2012A model). The data were acquired by the NSF-funded Center for Airborne Laser Mapping (NCALM) over the University of Houston campus and the neighbouring urban area (
Two separate processing steps were considered for the two different data source. 1) For fused data, we firstly fusion LiDAR and hyperspectral data using layer stacking, and employ the following technique:
• Reduce data dimensionality using Minimum Noise Fraction (MNF) transform.
• Select spectral endmember candidates to reduce spatial data dimensionality using Pure Purity Index (PPI).
• Evaluate linearity and select endmembers using n-D visualizer.
• Mapping endmember distribution using Spectral Angel Mapper (SAM) technique.
• Majority/minority analysis.
• Raster-vector conversion.
2) Urban tree heights then extracted from LiDAR data using urban vegetation class vector data as boundary, furthermore only trees with height more than equal to 1 meter and less than 30 meter extracted. We performed heights filtering technique, all return pulses were used, with canopy spacing 1 m, linear interpolation used for this step and finally geo-processing analysis employed to clip only the urban vegetation.
Human involvement appears in order to select the endmembers using a n-D visualizer, therefore this method called semi-automatics technique. Diagram flow of data processing shown in
We fused hyperspectral and LiDAR data at single resolutions to assess differences in classification accuracy with only hyperspectral (non-fusion) data. We then subset the three visible bands of hyperspectral-only data (B1, 502.5 nm; B2, 602.6 nm; B3, 697.9 nm) to produce a composite image near-natural color of the study area with an Red-Green-Blue (RGB) combination B1-B2-B3 for visual interpretation using superimposed technique. We subset the study area to only 100 ha enclosed area.
The calculation of the forward MNF transform occurs in two steps. The first step calculates the noise statistics based on a shift difference method. The second step performs the two principal components transformations. The MNF outputs data that has unit variance, isotropic noise
and has decorrelated data sorted by descending variance, the MNF transform will run out of information to process before it runs out of bands, leaving the final MNF output bands as noise-only proof of the hyperspectral nature data [
The Pixel Purity Index (PPI) is used to find the most “spectrally pure” or extreme, pixels in multispectral and hyperspectral data [
The PPI process exploits convex geometry concepts in the n-dimensional data space of the MNF-processed hyperspectral data. The purest pixels still must be on the extreme corners or edges of the data cloud. While n-dimensional data may be hard to visualize and imagine, the PPI process is relatively insensitive to dimensionality and acts as an n-dimensional “rock tumbler” or “clothes dryer”, tumbling the n-dimensional cloud of data points and counting how many times each pixel is “hit” in this tumbling process. The purer the pixel, the more convex the data cloud is at that location and as a result it will be hit more often and receive a PPI score higher than a lesspure pixel. This convexity concept is based on the assumption of hyperspectral over determinacy [
n-dimensional visualizer provides an interactive tool for finding endmembers by locating and clustering the purest pixels in n-dimensional space. Spectra can be thought of as points in an n-dimensional scatter plot, where “n” is the number of MNF bands or dimensions. The coordinates of each point in n-space consist of “n” values that are simply the spectral radiance or reflectance values in each band for a given pixel [
The Spectral Angle Mapper (SAM) matches image spectra to reference spectra in n-dimensions. SAM compares the angle between the endmember spectrum (considered as a n-dimensional vector, where n is the number of bands) and each pixel vector in n-dimensional space. Smaller angles represent closer matches to the reference spectrum. SAM produces a classified image based on the SAM Maximum Angle Threshold. Decreasing this threshold usually results in fewer matching pixels. Increasing this threshold may result in a more spatially coherent image, however, the overall pixel matches will not be as good as for the lower threshold [
It can be seen from n-D visualizer endmember spectra (
The area of urban vegetation is ~32 ha with height variations ranging from ~12 m to ~30 m. The spatial distribution of higher urban trees is in residential areas, while the lower urban trees spread along the city streets.
The small beam may be completely absorbed by the canopy before it reaches the floor and/or may miss the tree tops causing an underestimation of tree height [
The basis for LiDAR measurement is that a laser pulse is sent out from the sensor and the leading edge of the returned signal trips a response for a time measurement. The trailing edge of the response is also used to trip a second return time. These are referred to as the “first” and “last” returns. If the first return happens to be associated with a tree canopy top and the last return the underlying ground, then this single signal can be used to provide a measurement of tree height.
Complex backscatter signal (waveform) LiDAR systems typically have a much larger footprint that discrete return systems, being of the order of 10s of metres. This is fundamentally for signal-to-noise reasons: the quantity
of backscattered energy in a small field of view is low. The energy received per unit time bin is clearly even smaller, so the sensor technologies need to be capable of measuring very low signal levels, very quickly. The LiteMapper-5600 system quotes a waveform sampling interval of 1 ns, giving a multi-target resolution (related to bin size) of better than 0.6 m [
This concept can be found in
The magnitude of the ground reflection signal is gen-
erally high in urban environment [
This technique allows us to generate the 3D images of the urban vegetation canopy, providing information on tree heights with high accuracy using fused hyperspectral and LiDAR data.
We have shown that this study successfully examine the advantage of fused hyperspectral and LiDAR data for mapping and monitoring urban tree heights in convenience yields. This semi-automatics technique gives an efficient framework to be applied for mapping and monitoring urban trees in other cities around the world. However, in this study topography variable is neglected, therefore this method only suitable for the relatively flat urban environment.
The author would like to thank the Hyperspectral Image Analysis group and the NSF Funded Center for Airborne Laser Mapping (NCALM) at the University of Houston for providing the data sets used in this study, and the IEEE GRSS Data Fusion Technical Committee.
DSM (Digital Surface Model)
LiDAR (Light Detection and Ranging)
LULC (Land Use Land Cover)
MNF (Minimum Noise Fraction)
NCALM (NSF-Center for Airborne Laser Mapping)
NSF (National Science Foundation)
PPI (Pure Purity Index)
RGB (Red-Green-Blue)
SAM (Spectral Angel Mapper)
SRU (Spectral Radiance Units)