Mapping of Coastal Wetlands via Hyperspectral AVIRIS Data M. M. Crawford (1), A. L Neuenschwander (1), M. J. Provancha (2) (1) Center for Space Research, University of Texas at Austin (2) Dynamac Corporation, Kennedy Space Center, Florida
CLASSIFICATION OF WETLANDS USING AVIRIS DATA Project Goal: Investigate the potential of AVIRIS data for wetland vegetation classification at the Kennedy Space Center in Florida. Monitor the effects of various marsh management strategies by mapping the vegetation distribution and its change over time.
OPTICAL IMAGERY OF KENNEDY SPACE CENTER 1996 AVIRIS (Bands 49, 29, 20) of western shore of Kennedy Space Center 1989 Landsat TM (Bands 4,3,2) of Kennedy Space Center
GIS MAPS OF TEST SITE Map of Impoundments Vegetation map derived from 89 TM and aerial photography
AVIRIS IMAGERY Airborne Visible/Infrared Imaging Spectrometer flown by NASA’s Jet Propulsion Laboratory 224 Bands with 10 nm wavebands Measures visible and near infrared reflected energy (400 - 2500 nm) Airborne Platform flown 20 km above surface Highly calibrated instrument
PRE-PROCESSING OF AVIRIS DATA Atmospheric correction using ATREM* (developed by University of Colorado, CSES) Columnar Water Vapor removed from AVIRIS data. Spectral Transect of “Raw” and “Corrected” AVIRIS data.
FEATURE EXTRACTION Principal Components Analysis Minimum Noise Fraction (MNF) Transformation Orthogonal bands ordered by noise content Developed specifically for analysis of multi-band remotely sensed data Decision Boundary Feature Extraction
PRINCIPAL COMPONENTS ANALYSIS PC 1 PC 2 PC 3 PC 4 PC 5 PC 6
MINIMUM NOISE FRACTION (MNF) TRANSFORMATION MNF Band 1 MNF Band 2 MNF Band 3 MNF Band 4 MNF Band 5 MNF Band 6
CLASSIFICATION ALGORITHMS INVESTIGATED Pixel-Based Gaussian Maximum Likelihood Neural Network: (Multi-Layered Perceptron with one hidden layer and Scaled Conjugate Gradient Training algorithm) Canonical Analysis Region-Based Gaussian Markov Random Field
HIERARCHICAL METHODOLOGY Level 1 Level 2 Level 3
CLASSIFIER INPUTS Directly compare results of input combinations using a variety of classification algorithms 5 corrected AVIRIS Bands First 13 eigenvectors of MNF transformation 8 eigenvectors from Principal Components Analysis 32 upland and 11 wetland features extracted from Decision Boundary Feature Extraction (DBFE) algorithm 7 Upland and 5 Wetland features extracted from Canonical Analysis (CA)
PIXEL-BASED CLASSIFIER RESULTS NN5 MLC5 NN-MNF13 MLC-DBFE MLC-CA Vegetation Type Classifier Type Scrub Willow Marsh CP Hammock CP/Oak Hammock Slash Pine Oak Hammock Hardwood Swamp Uplands Total Graminoid Marsh Spartina Bakerii Marsh Typha Marsh Salt Marsh Mud Flats Wetlands Total 95.3 84.8 97.2 81.3 80.2 93.8 90.5 96.3 87.2 88.5 87.1 87.9 88.2 82.8 93.8 63.1 63.1 91.7 73.8 77.4 62.3 67.1 83.0 53.4 94.4 46.0 47.2 90.8 98.3 83.8 88.6 94.3 100.0 69.5 89.5 76.6 76.4 92.5 78.0 86.8 74.8 74.2 98.6 80.3 79.1 87.3 89.6 90.8 77.9 83.5 83.6 90.8 94.8 75.7 82.2 97.1 96.9 99.3 88.1 87.4 92.8 73.6 83.8 79.9 82.7 87.1 85.0 93.5 80.4 83.0
CONTEXTUAL CLASSIFIER RESULTS MRF-PC8 MRF-MNF13 Classifier Type 93.3 93.4 90.1 93.0 89.5 93.4 74.6 83.7 94.4 78.3 96.9 92.6 96.2 98.1 90.7 90.3 81.9 82.6 91.5 87.3 95.0 95.0 98.3 95.2 79.5 79.7 89.2 87.9 Vegetation Type Scrub Willow Marsh CP Hammock CP/Oak Hammock Slash Pine Oak Hammock Hardwood Swamp Uplands Total Graminoid Marsh Spartina Bakerii Marsh Typha Marsh Salt Marsh Mud Flats Wetlands Total
CLASSIFICATION RESULTS Gaussian MRF using MNF transformation input data. Uplands: 90.3 % Wetlands: 87.9%
CLASSIFICATION RESULTS Neural Net using MNF transformation input data. Uplands: 92.5 % Wetlands: 93.5 %
CLASSIFICATION RESULTS Gaussian MLC using 5 original AVIRIS bands as input data. Uplands: 76.4 % Wetlands: 85.0 %
CONCLUSION Hierarchical classification methodology was utilized MNF Bands used as input to classifier yielded best results Gaussian Markov Random Field contextual model classifier yielded best results Hyperspectral imagery is effective for classification of coastal wetlands