Page 1 of 1

Target Identification in multispectral images using in-situ hyperspectral reflectance

Atma Bharathi M*#, Suresh Kumar C*, Dr. S. Kalliappan**

*Institute of Remote Sensing, College of Engineering Guindy, Anna University Chennai.

#Corresponding author: atmabharathi@gmail.com

**Professor, Former director, Institute of Remote Sensing, CEG, AU Chennai.




Abstract
Targets are best identified on hyperspectral images due to their increased spectral resolution. But, due to the high cost and inherent difficulties, acquisition of those images becomes a rarity. Hence we researched on the capacity of commercially available multispectral images to identify targets when analyzed with in-situ hyperspectral reflectance.

A QuickBird image of Anna University Chennai campus captured two years before field data collection was used. The image was radiometrically corrected by the data provider and it was calibrated using Empirical Line Calibration (ELC) to reflectance, removing residual solar irradiance and atmospheric path radiance effects. ELC was performed first using only bright and dark pixels, then using bright, dark and intermediate pixels and the results were compared.

In-situ reflectance was collected using ASD Field Spec Hand Held hyperspectral radiometer in the region (325nm to 1075nm). It was then resampled by averaging intermediate points to match the satellite sensor bandwidth. Spectral Angle Mapper algorithm was used to identify targets in the image by comparing the angular deviation between the image spectrum and in-situ spectrum for each target at every pixel.

The procedure was performed for original, pan-sharpened images with and without ELC correction and the results were observed. We found that pan-sharpened and original images gave best results when used after ELC calibration with bright, dark and intermediate values scaled to reflectance. The results improved further when we used the optimal band centers we suggest. The limited, relatively broad spectral bands of QuickBird were capable of classifying all endmembers to level-2 accuracy. While they could pick up soil, different roof materials, roads and grass to 70% accuracy, identification of individual species of plants was comparatively poorer, thus depicting its limited level 3 classification capacity.

Introduction
Hyperspectral images with hundreds of, narrowly spaced spectral bands have very high spectral resolution. But due to their reduced radiometric resolution, hyperspectral sensors are flown mostly on airborne platforms. Thus due to the inherent difficulties in planning a flight; cost of acquiring such images, hyperspectral remote sensing is still in research stages.

On the other hand, multispectral images have high spatial and radiometric resolution even while they are on satellite platforms which enable economic and quick acquisition. But the reduced spectral resolution leaves much of their information content to go untapped.

In our project we research on the efficiency of a hybrid approach which uses the in-situ hyperspectral reflectance of specific targets to identify them on a multispectral image; thus bridging the advantages of both types of remote sensing.

The objectives of this project were to make a hyperspectral library of field collected reflectance covering the spectral bandwidths of high spatial resolution multispectral satellite sensors and to correct the multispectral image for atmospheric effects and to identify targets in it with the library created.

Data used
A Quick bird multispectral imagery imaged during February 2005 was used. The satellite has sensors covering the wavelengths 450-520 nm in blue, 530-590 nm in green, 630-690 nm in red and 770-900 nm in near infra-red regions. It had a spatial resolution of 2.5 m.

For in-situ field reflectance, an ASD Field Spec Hand Held Hyperspectral radiometer that works in 325 to 1075nm region with a spectral resolution of 1.6nm in visible and 3nm in NIR region was used. Sampling was performed during the same season but in the year 2007.

Methodology

Field reflectance collection
The study area was urban and was divided into four possible urban spectral endmembers – roads, building roof, soil and vegetation. A total of 31 samples of which 20 were different tree canopies were recorded. The samples were recorded between 10:30 to 11:45 AM. The scaled absolute reflectance of the samples was derived using the formula

Rabsref = (Isample / Iref ) Rref
Where Rabsref – absolute reflectance,
Isample – amount of light reflected from the sample,
Iref – amount of light reflected from calibrated barium sulphate white reference,
Rref – reflectance of the reference (here 1).

Observation was made under clear skies and the spectrum was constantly checked for artifacts. Errors which generally occur during field spectra collection (from Gu et al. 1992) were minimized by the following methods.
  • Error due to diffuse irradiation (skylight) - This was minimized by avoiding days when the sky was overcast
  • Error due to the reflectance panel - Calibrated barium sulphate white reference was used. It had Directional / isotropic-hemispherical reflectance factor when the panel is being uniformly illuminated and viewed in a specified angle.
  • Error due to non simultaneous sampling of target and reference panel - Ideally the reflected flux from the target and panel should be measured simultaneously. Since only one instrument was used this was not possible and the technique adopted is called Single beam method. As per the requirement the solar disc was free of cloud between reference and target measurements.
  • Error due to time delay between successive samples - Measurements were made near solar noon when the solar geometry is changing least and when the errors due to the angular response of the reflectance panel are at a minimum.
  • Errors introduced by changes in the target surface due to heliotropism in plants or disturbance due to wind were minimized by a time averaged measurement.

Figure 1 Absolute reflectance of endmembers. The reflectance of white reference is 1.


Creation of spectral library in ENVI
The radiometer spectra had artifacts in UV (325 – 400nm) region and at the 1000 to 1075nm region. Hence this region was cropped off and a library from 420 to 970nm for 31 common urban endmembers was created in ENVI software. Botanical names were given to identify plant species.

The library was resampled by averaging intermediate values to match QuickBird bandwidths; values between two bands were considered as zero during averaging - just the way the satellite sensor does. The figure 2 shows original and resampled spectra with the Full Width at Half Maximum (FWHM) at each of the bands.


Figure 2 Original spectra in dotted line and resampled field spectra in continuous line.


Atmospheric correction
Data collected from satellite sensors must be converted from raw radiance values to atmospherically corrected reflectance values to allow spectra to be compared with reference spectra in spectral libraries (Kruse 1994). Atmospheric correction allows comparison of field and satellite data even when they are temporally different (Song et al. 2001). However, radiative transfer models produce artifacts in the image spectra which have to be corrected using scene based methods like Empirical Line Calibration (ELC) (Clark et al. (1995) and Geotz et al. (1998)).

Empirical Line Calibration (ELC)


Figure 3 Empirical line calibration using two targets


Two targets (light and dark) with known reflectance (R) and at-sensor radiance (L) are joined by a straight line with slope s and intercept a. The reflectance for any at-sensor radiance can be computed from R = s (L-a). The term „a’ represents atmospheric radiance. This equation is computed for all four spectral bands. (Smith and Milton 1999).

The darkest and brightest pixels in the image were found using image statistics and 2D scatter plot for each band. When the pixels in scatter plot are selected, (as shown in fig 4) the corresponding pixels in the image were highlighted and their image spectra were extracted for ELC.


Figure 4 Scatter plot used to select the darkest and brightest pixels


The objective of ELC is to determine the slope 's' (multiplicative factor / gain) and intercept 'a' (additive factor / offset) from the bright and dark pixels and apply the same to rest of the pixels to determine their reflectance. ELC forces image spectra to match field reflectance spectra (Cone et al. 1987) and thus normalizes both the spectra for further analysis.

When we used spectra of intermediate pixels along with brightest and darkest pixels, the effects of ELC improved dynamically. Similar effects were reported by Goetz et al. (1997). By applying ELC to radiometrically corrected image, the resultant solar irradiance and atmospheric path radiance effects are removed (R.J. Aspinall et al 2001) and the resulting spectra is comparable to most field and laboratory spectra (Smith and Milton 1999).

Target Identification
Among the several spectral analysis algorithms (like Binary encoding, Continuum removal, Spectral Feature Fitting, Unmixing methods), Spectral Angle Mapper (SAM) best suits our work since it measures only the angular deviation between two spectra making it insensitive to illumination effects.

SAM is an automated method for comparing image spectra to a spectral library (Kruse et al., 1993a). It works on images with apparent reflectance and determines the similarity between two spectra by calculating the “spectral angle” between them, treating them as vectors in a space with dimensionality equal to the number of bands (nb). A smaller angle means a closer match between the two spectra and the pixel is identified as the field spectrum. Because it uses only the “direction” of the spectra, and not their “length,” the method is insensitive to the unknown gain factor, and all possible illuminations are treated equally. The same target would have lesser reflectance in satellite imagery compared to field spectra due to attenuation effects. Thus SAM ignores this anomaly by considering only the angle between the spectra and not their actual reflectance values.

SAM determines the similarity of an unknown spectrum t to a reference spectrum r, by applying the following equation (CSES, 1992)


which also can be written as:


where nb equals the number of bands in the image.

For each reference spectrum chosen in the library, a spectral angle a is determined for every image spectrum (pixel). This angle in radians is assigned to the corresponding pixel in the output SAM image. The angles are subjected to thresholds to determine their class.

Results I
SAM was used to identify targets in multispectral images of QB, both before and after ELC calibration. Since SAM measures only angle between spectra, it can work on images with radiance expressed in DN also. After several trials, we concluded that 0.1 radians tolerance provided best results. Results showed that
  • Image without ELC correction (Radiance in DN values) picked up only one target. It had high commission error.
  • Image calibrated using only bright and dark values for ELC contained residual artifacts. Image statistics showed a maximum reflectance of 2 and minimum of -1.8. This image showed a drop in efficiency, all targets were wrongly identified.
  • Images that were calibrated using bright, dark, and intermediate values for ELC identified much more targets. Roads, roofs, soil were picked up while vegetation identification was poor.

Figure 5 Target identification results 1


The overall target identification capacity was not satisfactory. During this process, we discovered that the present band center configuration of the resampled spectrum was inefficient in picking up targets that had only subtle differences in their spectrum. This difficulty was more pronounced in vegetation species identification. When all the vegetation spectra were plotted, the angular differences between them were too small that they would fall within the tolerance limits of the SAM classifier. On the other hand making the tolerance more stringent would increase the omission errors.

This paradox was resolved when we slightly shifted the band centers to positions in the spectra where the slope of the spectra is more enhanced. The following figure explains this concept.


Figure 6 Optimal band centers


The continuous curve is the hyperspectral field spectra. The dotted line is the spectrum resampled to present QB band centers. The continuous line is the spectrum resampled to optimal band centers we have evolved. It can be seen that the continuous line matches more closely to real spectra. The following table shows actual and optimal band centers.


Figure 7 Comparing actual and optimal band centers of quick bird satellite


Results II
Next SAM was applied by resampling the library to the new band centers before and after ELC and the results were observed as follows.
  • Images without ELC correction identified all targets except vegetation.
  • Images calibrated through ELC with only bright and dark values showed much lesser efficiency than images without calibration.
  • Images calibrated with bright, dark, intermediate values for ELC showed a phenomenal increase in accuracy. 60% of the targets were picked up.

Figure 8 Classification results 2


Discussion
On the whole, the QuickBird image showed acceptable capability to pick up non-vegetation targets. Plant species had only subtle variation even when recorded by the hyperspectral radiometer. Thus when the spectra were resampled to coarse spectral resolution of QuickBird, species differentiation became difficult.

Further, the temporal shift of 2 years between the image acquisition and field data collection should have played a major role in modification of target characteristics. Plant canopy should have grown, their architecture could have changed. Further we found the tight and loose soil blatantly missing in the satellite imagery. Also red tile roof could have had additional algal growth owing to heavy downpour during the earlier two monsoons.

One situation where the illumination indifference character of SAM becomes a disadvantage is when two bright and dark pixels have the same spectral curves as shown below.


Figure 9 SAM fails to distinguish between a dark and bright pixel since they have a similar curve


SAM fails to distinguish between the two targets and classifies them as same. In reality they were collected from a swimming pool floor and other on tar roof.

Further, the satellite – nadir – solar geometry could not be recreated since the satellite observed the scene off-nadir.


Figure 10 Satellite, Earth, Sun geometry


The graphics show the top view, front view and perspective view of satellite's off nadir view.

While identifying trees, the canopy center had the correct match while the periphery was identified for other species. This throws light on the canopy architecture and the effects of adjacency on the outer pixels, while the center pixels remained comparatively unmixed.

Conclusion
The study reveals that the ability of multispectral images to identify targets is fair. ELC with intermediate spectra is required for best normalization and results, but this hinders the automation of target identification which would have otherwise required the user to just feed the imagery and spectral library.

The study can be extended by sampling all the urban materials to make a complete library. This way multispectral image of urban areas can be accurately classified up to levels 2, 3. To a certain level, species of vegetation can also be identified.

References
  • Boardman JW, (1989), “Inversion of imaging spectrometry data using singular value decomposition”, Proceedings, IGARSS?89, 12th Canadian Symposium on Remote Sensing, 4:2069–2072
  • Clark RN, King TVV, Gorelick NS (1987) “Automatic continuum analysis of reflectance spectra”, Proceedings, 3rd AIS workshop, 2-4 June, 1987, JPL Publication 87-30, Jet Propulsion Laboratory, Pasadena, California, p 138–142
  • Clark RN, Gallagher AJ, Swayze GA (1990), “Material absorption band depth mapping of imaging spectrometer data using the complete band shape least-squares algorithm simultaneously fit to multiple spectral features from multiple materials”. Proceedings of the 3rd Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) Workshop, JPL Publication 90-54, pp 176–186.
  • Clark RN, Swayze GA, Gallagher A, Gorelick N, Kruse FA (1991), “Mapping with imaging spectrometer data using the complete band shape least-squares algorithm simultaneously fit to multiple spectral features from multiple materials”. Proceedings, 3rd Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) Workshop, JPL Publication 91-28, pp 2–3
  • Conel JE, Green RO, Vane G, Bruegge CJ, Alley RE, Curtiss B (1987), “Airborne Imaging Spectrometer-2: Radiometric spectral characteristics and comparison of ways to compensate for the atmosphere”. Proceedings SPIE, 834, pp 140–157
  • Gary A. Shaw and Hsiao-hua K. Burke, MIT, “Spectral Imaging for remote sensing”.
  • Goetz AFH, Heidebrecht KB, Kindell B, Boardman JW (1998), “Using ground spectral irradiance for model correction of AVIRIS data”. AVIRIS 1998 Proceedings, JPL, California. 10 pp (http://makalu.jpl.nasa.gov/docs/workshops/98_docs/22.pdf)
  • Gu, X, F, Guyot, G. and Vebrugghe. M 1992, “Evaluation of measurement errors in ground surface reflectance for satellite calibration”, International Journal of Remote Sensing, 13, 2531 – 2546.
  • Kruse FA, Lefkoff AB, Boardman JB, Heidebrecht KB, Shapiro AT, Barloon PJ, Goetz AFH (1993b), “The spectral image processing system (SIPS)–Interactive visualization and analysis of imaging spectrometer data”. Remote Sensing of Environment 44:145–163
  • Mazer AS, Martin M, Lee M, Solomon JE (1988), “Image processing software for imaging spectrometry data analysis”. Remote Sensing of Environment 24:201–210
  • Peg Shippert, Ph.D. Earth Science Applications Specialist Research Systems, Inc. “Introduction to Hyperspectral Image Analysis”.
  • Randall B. Smith, Ph.D., MicroImages, Inc., “Introduction to hyperspectral imaging”.
  • Smith GM and Milton EJ. 1999, “The use of the Empirical line method to calibrate remote sensing data to reflectance”. International Journal of Remote Sensing 20 2653 – 2662.
  • Song C, Woodcock, CE Seto, K.C. Lenny, M.P. and Macomber, S.A. 2001. “Classification and change detection using Landsat TM data, when and how to correct atmospheric effects?” Remote Sensing of the Environment 75, 230-244.
  • USGS Spectroscopy lab, Surface Reflectance Calibration of Terrestrial Imaging Spectroscopy Data