Hyperspectral Environmental
Monitoring of Waste Disposal Areas
Jason Hamel
Advisor: Rolando Raqueño
Digital Imaging and Remote Sensing Laboratory
Chester F. Carlson Center for Imaging Science
Rochester Institute of Technology
Rochester, NY
Digital Imaging and Remote Sensing Laboratory
Overview
• Background
• Procedure/Results
»Spectra
»Classification
• Conclusions
Digital Imaging and Remote Sensing Laboratory
What are Landfills?
• Very common waste management
technique
• They do not separate toxic wastes from
the environment
• A water resistant clay cap is placed over
the landfill slows the spread of chemicals
Digital Imaging and Remote Sensing Laboratory
Clay Caps
Diagram of the material layers in the 2 major clay
cap technologies [BGC.pdf for DOE’s SRS site]
Digital Imaging and Remote Sensing Laboratory
Clay Cap Technology
• Caps are designed to last 40 years
• Replace with new technology that actually
deals with the waste
• This lack of solution has given chemicals
time to leach into the environment
• These problem areas must be found
Digital Imaging and Remote Sensing Laboratory
Why Look at Landfills?
• Currently, possible dangerous sites are
manually sampled and processed in a lab
• This can be time and money consuming
for larger sites or a large number of sites
• Chemicals are often dangerous even at
very low concentrations
• Remote sensing with new hyperspectral
detectors may provide and economic
alternative
Digital Imaging and Remote Sensing Laboratory
Example of Expected Imagery
Hyperspectral
AVIRIS scene with
224 bands
SRS site
Digital Imaging and Remote Sensing Laboratory
Purpose of this Research
• Low concentrations make it very difficult to
directly detect a chemical’s spectral signature
• Determine if new hyperspectral sensors collect
enough information to identify materials
• Determine the detectability of specific
secondary spectral effects of leachates (e.g.):
» Vegetation health
» Soil water moisture
• Determine if atmospheric correction is
necessary
Digital Imaging and Remote Sensing Laboratory
General Procedure
PROSPECT Real Soil
Material Classification
OSP SAM SSM Unmixing
Spectra VEG
Spectral S/U
Matching
Algorithms VEG/
Wavelength Soil
Soil/
Atmosphere, Soil
Mix
Detectors,
Spectra
and Noise
Digital Imaging and Remote Sensing Laboratory
Vegetation Spectra
• PROSPECT leaf model and software
• Two varied inputs:
» Chlorophyll concentration (mm/cm2)
» Equivalent water thickness (cm)
• Generated spectra
» Healthy leaf (high chlorophyll and water)
» Stressed leaf (low chlorophyll and water)
Digital Imaging and Remote Sensing Laboratory
Vegetation Spectra
Reflectance
Spectra of
Vegetation
Green : Healthy
Red : Stressed
Digital Imaging and Remote Sensing Laboratory
Soil Spectra
• Ground measurements taken with
spectrometer as soil dried
• Moisture in soil was not measured while
spectra was taken
• Relative labels given to various spectra
» Wet Soil
» Moist Soil
» Dry Soil
Digital Imaging and Remote Sensing Laboratory
Soil Spectra
Reflectance
Spectra of Soil
Brown : Dry
Orange : Moist
Black : Wet
Digital Imaging and Remote Sensing Laboratory
Reflectance Data Set
• The 5 basic vegetation and soil spectra
are mixed by:
where R1 and R2
are 2 basic spectra
• This creates 10 additional mixed spectra
• 15 spectra in final data set
Digital Imaging and Remote Sensing Laboratory
Atmosphere and Detector Effects
• Light reflecting off material propagates
through atmosphere
• Detector measures the radiance reaching
the detector at various narrow wavelength
regions called channels
• Detector electronics record input signal in
digital counts (DC)
• The hyperspectral AVIRIS detector has
224 channels from 400nm to 2500 nm
Digital Imaging and Remote Sensing Laboratory
Digital Count Spectral Data Set
• Radiance reaching the sensor, Lsen,
calculated from the Big Equation:
ES
Lu LD
T2 T1
R
• Radiance variables supplied by MODTRAN
Digital Imaging and Remote Sensing Laboratory
Detector Effects
• All spectra converted to AVIRIS
wavelength regions
• Lsen was multiplied at each wavelength by
an AVIRIS gain factor to calculate AVIRIS
DC’s
Digital Imaging and Remote Sensing Laboratory
AVIRIS Basic DC Spectra
Digital Imaging and Remote Sensing Laboratory
Realistic Data Set
• All detectors measure noise as well as
signal
• Standard gaussian noise with standard
deviation of 1 added to DC spectra (not
representative AVIRIS noise value)
• Noisy sensor radiance determined
• Noisy reflectance spectra calculated by
removing atmosphere effects
Digital Imaging and Remote Sensing Laboratory
Noisy Basic Reflectance Spectra
Digital Imaging and Remote Sensing Laboratory
Classification
• 6 classification algorithms used:
» Linear Spectral Unmixing (ENVI)
» Orthogonal Subspace Projection (Coded)
» Spectral Angle Mapper (ENVI)
» Minimum Distance (ENVI)
» Binary Encoding (ENVI)
» Spectral Signature Matching (Coded)
• The 5 basic vegetation and soil spectra
were used as endmembers
• Reflectance endmembers converted to DC
before classifying DC spectra
Digital Imaging and Remote Sensing Laboratory
Classification Algorithms
• Linear Spectral Unmixing (LSU)
» Generates maps of the fraction of each endmember in a pixel
• Orthogonal Subspace Projection (OSP)
» Suppresses background signatures and generates fraction maps
like the LSU algorithm
• Spectral Angle Mapper (SAM)
» Treats a spectrum like a vector; Finds angle between spectra
• Minimum Distance (MD)
» A simple Gaussian Maximum Likelihood algorithm that does not
use class probabilities
• Binary Encoding (BE) and Spectral Signature Matching
(SSM)
» Bit compare simple binary codes calculated from spectra
Digital Imaging and Remote Sensing Laboratory
Classification Results
• The LSU and OSP fraction maps allow for
the calculation of sum of squared error:
i = endmember
j = wavelength
Sum of Square Errors
Classifier Ground Sensor DC Retrieved
Reflectance with Atmosphere Reflectance
LSU 4.81e-11 0.239 0.032
OSP 2.32e-6 0.237 0.038
Digital Imaging and Remote Sensing Laboratory
Classification Results
• The SAM, MD, BE, and SSM algorithms were not
designed to classify mixed pixels
• Accuracy is the correct identification of one of
the fractions in a pixel
Percent Accuracy
Classifier Ground Sensor DC Retrieved
Reflectance with Atmosphere Reflectance
SAM 66.67% 40.00% 66.67%
MD 66.67% 80.00% 66.67%
BE 86.67% 66.67% 86.67%
SSM 93.33% 80.00% 93.33%
Digital Imaging and Remote Sensing Laboratory
Conclusions
• Atmosphere degrades performance of
most of the classification algorithms
studied
• Removal of the atmosphere is
recommended
• The LSU and OSP fraction maps are more
useful
» Provide very accurate material identification
without a large spectral library
» Detects not just the material, but the amount of
material in a given pixel
Digital Imaging and Remote Sensing Laboratory
Follow Up Work
• There are many areas to expand on this
research
» More realistic sensor noise
» Additional levels of vegetation health
» Broader range of atmospheres
» Incorporate background cloud effects
» Create a greater variety of mixed pixels
–Different percentages
–More than 2 materials
» Identification of actual secondary spectral effects
of leachates
Digital Imaging and Remote Sensing Laboratory