Rapid Prototyping of Hyperspectral Image Analysis Algorithms for Improved Invasive Species Decision
Support Tools
FINAL REPORT NASA GRANT # DONNS06AA65D USMMRCSSC1216200565D 06040327
SUBMITTED TO: TED MASON, NASA STENNIS SPACE CENTER
SUBMITTED BY: LORI MANN BRUCE, PH.D. GEORESOURCES INSTITUTE
MISSISSIPPI STATE UNIVERSITY BOX 9627
MISSISSIPPI STATE, MS 39762-9627 PHONE – 662-325-8430
EMAIL – [email protected]
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 2/15
Introduction Nonnative invasive species adversely impact ecosystems, causing loss of native plant
diversity, species extinction, and impairment of wildlife habitats. As a result, over the past
decade federal and state agencies and nongovernmental organizations have begun to work
more closely together to address the management of invasive species. In the 2005 fiscal year,
approximately $500M was budgeted by U.S. Federal Agencies for the management of
invasive species, where the primary funding included the following: DOD-$15M, USDA-
$470M, and DOI-$11M [Simberloff 2005]. Despite extensive expenditures, most of the
methods used to detect and quantify the distribution of these invaders are ad hoc, at best.
Likewise, decisions on the type of management techniques to be used or evaluation of the
success of these methods are typically non-systematic. More efficient methods to detect or
predict the occurrence of these species, as well as the incorporation of this knowledge into
decision support systems, are greatly needed.
The invasive species known as salt cedar or tamarisk (Tamarix ramosissima) is a
particular problem in the U.S.’s desert southwest, where it is displacing the native
cottonwood, willow, and other native plants. Tamarisk shrubs, or trees, are extremely
competitive against native vegetation because they aggressively consume the water supply.
Since tamarisk can re-grow from root crown buds, even after burning, the current
management practices for tamarisk involve combinations of chemical, mechanical, and
biological techniques. Thus, detection of tamarisk when it is in it earliest growing stages,
through the use of remote sensing, could greatly reduce the cost associated with this invasive
species.
In this rapid prototyping capabilities (RPC) experiment, recently developed analysis
techniques for hyperspectral imagery were prototyped for inclusion in the National Invasive
Species Forecasting System (NISFS). The new decision methodologies were tested using
hyperspectral data obtained with a handheld spectroradiometer and hyperspectral imagery
obtained via NASA’s HYPERION sensor. The results of this RPC experiment clearly
demonstrated the viability of applying remotely sensed imagery, particularly hyperspectral
such as that of HYPERION, to the problem of invasive species detection, as well as the image
analysis methods’ potential for improving the existing NISFS decision support system.
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 3/15
Figure 1: Tamarisk stand in Colorado. Native plants are unable to penetrate the thick stands of tamarisk [Photograph: Tim Carlson].
Figure 2: Removal of dead tamarisk using fire at the Bosque del Apache NWR, NM [“Cost Components for Non-Native Phreatophyte Control, November 2006”, Tamarisk Coalition]
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 4/15
Existing NISFS Decision Support System
In the three years prior to this RPC experiment, the PIs participated in the NASA invasive
species science team meetings for the development of the NISFS
[http://bp.gsfc.nasa.gov/isfs_people.html]. We presented our work at the science team
meetings, and these presentations are listed on the NISFS website
[http://bp.gsfc.nasa.gov/isfs_news.html]. As a result of our meetings with the science team,
we worked with members from the team to develop hyperspectral image analysis tools for
automated detection of invasive vegetation. Likewise, the PIs are beginning work on a task
force, as part of the science team, to develop a “library” of spectral-temporal profiles for
invasives. It is currently planned for the National Institute of Invasive Species Science
(NIISS) to serve as a clearinghouse, and the library will be developed in conjunction with the
USGS spectral library program. This is particularly of interest to the RPC experiment, since
the “library” could be potentially used as training data for future (non-tamarisk) applications
of the experiment’s resulting tools: hyperspectral image analysis algorithms for detecting
invasive species.
RPC Experiment: Methodologies
Hyperspectral Imagery Analysis and Automated Target Recognition Techniques
Over the past six years, the PIs have researched the use of remotely sensed multispectral
and hyperspectral reflectance data for the detection of invasive vegetative species [refs 1-15].
As a result, the PI has designed, implemented, and benchmarked various target detection
systems that utilize remotely sensed data. These systems have been designed to make
decisions based on a variety of remotely sensed data, including high spectral/spatial resolution
hyperspectral signatures (1000’s of spectral bands with spectral resolutions of approximately
3nm, such as those measured using ASD™ handheld devices), moderate spectral/spatial
resolution hyperspectral images (100’s of spectral bands, such as HYPERION imagery), and
low spectral/spatial resolution images (such as LANDSAT or MODIS imagery). These
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 5/15
algorithms (discrete analysis methods) include hyperspectral exploitation techniques such as
the following dimensionality reduction methods:
• Best spectral band selection (i.e. stepwise-LDA band selection)
• Optimized spectral band grouping
• Stepwise PCA component selection
The PIs have extensive experience with combining these recently-developed methods with
conventional feature reduction methods, such as Fisher’s LDA, and conventional classifiers,
such as nearest neighbor and maximum likelihood classifiers. The end result is an end-to-end
automated target recognition (ATR) system for detecting invasive species. The outputs of
these systems can be invasive prediction maps, as well as quantitative accuracy assessments
like confusion matrices, user accuracies, and producer accuracies.
Figure 3. Simplified block diagram of discrete analysis pattern recognition techniques applied to hyperspectral imagery.
Hyperspectral Data and Verification and Validation Techniques
Collaborators at Colorado State University (CSU) and USGS supplied the necessary
ground truth via field surveys. Collaborators at NASA supplied hyperspectral signatures from
handheld sensors, as well as ASTER co-registered HYPERION imagery of the field sites.
The ground truth and hyperspectral data were both collected in the Grand Staircase-Escalante
National Monument, in southern Utah. All model design, algorithm development,
experimental analysis, and verification and validation (V&V) were conducted by the MSU
researchers.
Dimensionality Reduction
Feature Optimization Classification
Pixel Label (Map)
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 6/15
Figure 4. Hackberry Canyon in the Grand Staircase-Escalante National Monument was used as the field study test sight for all experimental data collection (ground truth and HYPERION imagery).
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 7/15
Figure 5. Airborne and satellite imagery of field study test site, Hackberry Canyon in the Grand Staircase-Escalante National Monument.
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 8/15
July 2004
November 2004 December 2004
October 2004August 2004July 2004
November 2004 December 2004
October 2004August 2004
Figure 6. Temporal photos of tamarisk in Hackberry Canyon, Grand Staircase-Escalante National Monument. [Courtesy of Paul Evangelista].
Figure 7: HYPERION imagery color composite of Hackberry Canyon in the Grand Staircase-Escalante National Monument, with ground truth points overlaid. Red dots indicate ground truth points of tamarisk presence and blue dots indicate ground truth points of non-tamarisk.
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 9/15
RPC Experiment: Motivation
For all of the research endeavors described above, the PIs have developed numerous
advanced signal and image processing methodologies, as well a suite of associated software
modules. These methodologies have been extensively researched, benchmarked, and
published in refereed conferences and journals. However, prior to this RPC experiment, the
use of the prototype software modules has been primarily contained to in-house use at
Mississippi State University (MSU). This RPC experiment enabled
(i) a systematic verification and validation of the hyperspectral image analysis
algorithm using both handheld spectroradiometer data and satellite image data,
namely HYPERION imager;
(ii) a formal collaboration between remote sensing and invasive species researchers at
MSU (Lori Mann Bruce, et al), USGS (Tom Stohlgren, et al), CSU (Paul
Evangelista, et al), and NASA (Jeff Morisette, et al);
(iii) a direct comparison of invasive species mapping capabilities of the existing
decision support system (NISFS) and the proposed methods (automated analysis
of HYPERION imagery).
Experimental Results
The results of this research project clearly demonstrate the viability of applying
remotely sensing imagery, particularly hyperspectral such as that of HYPERION, to the
problem of invasive species detection. As expected, the results show that the tamarisk
detection accuracy depends on both spatial and spectral resolutions. The newly developed
image processing and pattern recognition methods resulted in accuracies as high as 90% for
discriminating tamarisk from native vegetation, such as cottonwoods and willows, when the
input hyperspectral sensors had a spectral resolution in the range of 30-100nm and a spatial
resolution that allowed for target abundances to be greater than 50%, i.e. mixed pixels had at
least 50% ground coverage of tamarisk. These results are very promising and clearly indicate
the power and practicality of hyperspectral remote sensing for invasive species detection.
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 10/15
350 550 750 950 1150 1350 1550 1750 1950 2150
0
0.2
0.4
0.6
0.8
1
1.2Tamaraisk (red)
Non-Tamarisk (blue, Cottonwood and Willow)
350 550 750 950 1150 1350 1550 1750 1950 2150
0
0.2
0.4
0.6
0.8
1
1.2Tamaraisk (red)
Non-Tamarisk (blue, Cottonwood and Willow)
Figure 8. High spectral and spatial resolution hyperspectral signatures of the target invasive vegetation (tamarisk) and nontarget non-invasive vegetation (cottonwood and willow). Signatures are preprocessed with waterband interpolation, truncation to 1650 bands, and normalization to [0,1].
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 11/15
Figure 9: Classification accuracies (overall accuracies of two class system: tamarisk
vs. non-tamarisk (willow, cottonwood, etc)) for varying spatial and spectral resolutions.
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 12/15
20 40 60 80 100 120 140 160 1800
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
band number20 40 60 80 100 120 140 160 180
0
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
20 40 60 80 100 120 140 160 1800
1000
2000
3000
4000
5000
6000
7000
8000
9000
10000
band number
Figure 10. HYPERION hyperspectral signatures of the target invasive vegetation (tamarisk - red) and nontarget non-invasive vegetation (cottonwood and willow - blue). Signatures raw data prior to any preprocessing (removal of water bands, normalization, etc).
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 13/15
Target Abundance
50
55
60
65
70
75
80
85
90
95
100
100% Target(no mixing)
80% Target in Mixed Pixels
50% Target in Mixed Pixels
20% Targetin Mixed Pixels
Ove
rall
Acc
urac
y (%
)
Maximum Likelihood Classifier, Leave-one-Out TestingMaximum Likelihood Classifier, 10 Partition Cross Validation Testing
Target Abundance
50
55
60
65
70
75
80
85
90
95
100
100% Target(no mixing)
80% Target in Mixed Pixels
50% Target in Mixed Pixels
20% Targetin Mixed Pixels
Ove
rall
Acc
urac
y (%
)
Maximum Likelihood Classifier, Leave-one-Out TestingMaximum Likelihood Classifier, 10 Partition Cross Validation Testing
Target Abundance
50
55
60
65
70
75
80
85
90
95
100
100% Target(no mixing)
80% Target in Mixed Pixels
50% Target in Mixed Pixels
20% Targetin Mixed Pixels
Ove
rall
Acc
urac
y (%
)
Maximum Likelihood Classifier, Leave-one-Out TestingMaximum Likelihood Classifier, 10 Partition Cross Validation Testing
Maximum Likelihood Classifier, Leave-one-Out TestingMaximum Likelihood Classifier, 10 Partition Cross Validation Testing
Figure 11. Classification accuracies (overall accuracies of two class system: tamarisk vs. non-tamarisk (willow, cottonwood, etc)) for HYPERION mixed pixels for varying target abundance.
Transition of Results to ISS
The USGS-NASA team was also awarded a “Earth Science REASoN – CAN” grant to
develop an Invasive Species Data Service (ISDS) to provide customized, easily accessible
data products and tools to support invasive species management and policy decision-making.
The PIs have begun discussions with the principal investigators of the ISDS project on how to
use their ISDS to integrate our newly developed tools into the NISFS, as part of a future
Integrated Systems Solutions (ISS) project.
Acknowledgments The researchers gratefully acknowledge our Colorado State University and USGS
collaborators, including Tom Stohlgren and Paul Evangelista, and our NASA collaborators,
including Jeff Morrisette and Steve Ma.
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 14/15
References [1] J.E. Ball, L.M. Bruce, and N.H. Younan, “Hyperspectral pixel unmixing via spectral
band selection and DC insensitive singular value decomposition,” IEEE Geoscience and Remote Sensing Letters, Vol. 4, No. 3, Jul. 2007.
[2] L.M. Bruce, A. Mathur, J.D. Byrd, “Denoising and Wavelet-Based Feature Extraction of MODIS Multi-Temporal Vegetation Signatures,” GIScience & Remote Sensing, vol. 43, pp. 170-180, 2006.
[3] J. Li and L.M. Bruce, “Wavelet-Based Feature Extraction for Improved Endmember Abundance Estimation in Linear Unmixing of Hyperspectral Signals,” IEEE Trans. Geoscience and Remote Sensing, vol. 42, no. 3, pp. 644-649, March 2004. (See Correction to “Wavelet-Based Feature Extraction for Improved Endmember Abundance Estimation in Linear Unmixing of Hyperspectral Signals”, IEEE Trans. Geoscience and Remote Sensing, vol. 42, no. 5, pp. 1122, May 2004.)
[4] S. Prasad, L.M. Bruce, “Limitations of Subspace LDA in Hyperspectral Target Recognition Applications,” Proc. IEEE Geoscience and Remote Sensing Symposium(IGARSS), Barcelona, Spain, July 2007.
[5] S. Prasad, L.M. Bruce, “Hyperspectral Feature Space Partitioning via Mutual Information for Data Fusion,” Proc. IEEE Geoscience and Remote Sensing Symposium(IGARSS), Barcelona, Spain, July 2007.
[6] T.R. West, S. Prasad, L.M. Bruce, “Multiclassifiers and Detection Fusion in the Wavelet Domain for Exploitation of Hyperspectral Data,” Proc. IEEE Geoscience and Remote Sensing Symposium(IGARSS), Barcelona, Spain, July 2007.
[7] J. Ball, T.R. West, S. Prasad, L.M. Bruce, “Level Set Hyperspectral Image Segmentation using Spectral Information Divergence (SID) based Best Band Selection,” Proc. IEEE Geoscience and Remote Sensing Symposium(IGARSS), Barcelona, Spain, July 2007.
[8] S. Prasad, L.M. Bruce, "Information Theoretic Partitioning and Confidence based Weight Assignment for Multi-Classifier Decision Level Fusion in Hyperspectral Target Recognition Applications," Proc. of the SPIE Defense and Security Symposium, Orlando, Florida, USA, April 2007.
[9] Mathur, L.M. Bruce, D.W. Johnson, W. Robles, J. Madsen, “Exploiting Hyperspectral Hypertemporal Imagery with Feature Clustering for Invasive Species Detection,” Proc. IEEE Geoscience and Remote Sensing Symposium(IGARSS), Denver, CO, August 2006.
[10] J. Ball, L.M. Bruce, “Level Set Hyperspectral Segmentation: Near-Optimal Speed Functions using Best Band Analysis and Scaled Spectral Angle Mapper,” Proc. IEEE Geoscience and Remote Sensing Symposium(IGARSS), Denver, CO, August 2006.
[11] Mathur, L.M. Bruce, D.W. Johnson, W. Robles, J. Madsen, “Automated Stepwise Selection of Hyperspectral Hypertemporal Features for Target Detection,” Proc. IEEE Geoscience and Remote Sensing Symposium(IGARSS), Denver, CO, August 2006.
NASA RPC Experiment Final Report, Lori Mann Bruce, Ph.D., Mississippi State University, page 15/15
[12] T.R. West, L.M. Bruce, “Detecting Invasive Species via Hyperspectral Imagery using Sequential Projection Pursuits,” Proc. IEEE Geoscience and Remote Sensing Symposium (IGARSS), Denver, CO, August 2006.
[13] Mathur, L.M. Bruce, “Identification of Pertinent Regions in Spectro-Temporal Maps for Vegetative Target Detection,” Proceedings of the American Society of Photogrammetry and Remote Sensing 2006 Annual Conference (ASPRS 2006), Reno, NV, May 2006.
[14] J. Ball, L.M. Bruce, “Accuracy Analysis of Hyperspectral Imagery Classification using Level Sets,” Proceedings of the American Society of Photogrammetry and Remote Sensing 2006 Annual Conference (ASPRS 2006), Reno, NV, May 2006.
[15] D.W. Johnson, L.M. Bruce, “Spatial and Spectral Resolution Effects on the Use of Remotely Sensed Data for Detection of Invasive Species,” Proc. IEEE Geoscience and Remote Sensing Symposium (IGARSS), Seoul, Korea, July 25-29, 2005.
Appendix A: Budget I. Salaries and Wages 64,070$
1a. Faculty Salaries - Lori Mann Bruce$108,542 annual (5% academic year) 4,070$
1b Research Associate - John Ball$60,000 annual @ 100% $60,000
II. Fringe Benefits 23,362$ 2a. 1a @ 29.72 1,210$ 2b 1b @ 29.72 17,832$ 2c Tuition @ $480.00/mo per student 4,320$
III. Equipment -$
IV. Supplies (Expendable Supplies) -$
IV. Travel 8,500$ 4a. In-state 1,000$ 4b. Out-of-State 7,500$
V. ADP Expenses -$ TOTAL DIRECT COSTS 95,932$
Top Related