Vegetation cover information was determined for the South Florida Everglades using unsupervised and artificial neural network (NN)-based supervised classification techniques applied to multisensor, multispectral, and RADAR imagery. The main objective was to develop and demonstrate an automated, supervised classification process for mapping vegetation type and density from moderate spatial resolution (30-m) Landsat thematic mapper (TM) multispectral imagery, integrated with higher spatial resolution synthetic aperature RADAR (SAR) data of the European Space Agency satellite.

Unique to this mapping requirement, the defined land cover classes can be mixtures of homogeneous cover types. For example, within the 30- by 30-m ground-projected instantaneous field of view (GIFOV) of the TM multispectral sensor, the cover for the low-density sawgrass class includes open water and sparse sawgrass. Consequently, the spectral signal corresponds to a mixture of ground covers instead of to a homogeneous area on the ground. Mixed pixels also occur when boundaries between homogeneous cover types are within the GIFOV; for example, the boundary between dense, homogeneous sawgrass and open water. The accurate classification of mixed pixels is difficult to obtain from multispectral imagery with the often used parametric maximum likelihood classification techniques applied on a per-pixel basis. Spectral unmixing techniques offer an alternative to the classification of mixed pixels, but there are other problems, such as the requirement for reference, or end-member spectra, for each class (Foody and others, 1997).

The application of multilayer NN techniques for the supervised classification of mixed pixels has been reported (for example, by Foody and others, 1997). The vegetation density mapping described here used a hybrid classification process that combined unsupervised, fuzzy clustering with supervised NN classification techniques. The overall process is thus NN-based classification and fusion of fuzzy-clustered, preprocessed, multisensor data. For this vegetation density mapping requirement, the individual NN outputs represent class (for example, vegetation type), while their output levels represent the fuzzy categories of low-, medium-, and high-density vegetation. With training, the NN learns to produce the three output levels that denote the respective density classes. This learned classification is optimal in the sense that the training process attempts to minimize the square error between the NN learned output level for class and the true level. It provides an automated method for mapping the image feature data into class; in this case, into the fuzzy (low, medium and high) -density classes. A class cover map results from hard thresholding the output values.

Input data for training were obtained by a two-step process. First, clusters were found by an iterative fuzzy clustering process (Fuzzy Learning Vector Quantization) (Bezdek, 1993) applied to the feature data derived from preprocessed multispectral (XS) and SAR imagery. The XS and SAR data were each clustered separately, using 10 and 8 clusters, respectively; simply more clusters than classes were used.

NN input values for each class-labeled feature vector used for training were the respective fuzzy memberships (a continuous value, proportional to the distance to the cluster) between the feature vector and all clusters. This may be contrasted to hard clustering, where the input feature data are associated only in a binary sense and with only one cluster. Information related to the other clusters is lost.

Preprocessing and feature extraction techniques were developed and applied to the TM and SAR data before clustering. For example, new techniques were developed for the resolution enhancement of TM 120-m thermal infrared data to improve classification results (Lemeshewsky, 1998, 2000). Also, a discrete wavelet transform (DWT)-based technique for the fusion of TM and higher spatial resolution panchromatic imagery (Lemeshewsky, 1999) was developed to produce several false-color 1:24000-scale photoimage maps. The SAR speckle noise was reduced using the reported (DWT)-based soft-thresholding technique of Donohoe and was implemented with shift-invariant DWT (Lemeshewsky, 1999). To improve the classification accuracy of TM mixed pixels, especially where the GIFOV contains boundaries between cover types, a reported (Wu and Schowengerdt, 1993) ‘partial restoration’ preprocessing technique was applied to the TM spectral band data, excluding the thermal band. The TM pixel feature data vector used in clustering included TM bands 2, 4, 5, and 7 after ‘partial restoration’ preprocessing, resolution-enhanced band 6 thermal data, and normalized difference vegetation index (NDVI) computed from partial restoration processed bands 3 and 4. Local mean and variance features were derived from the denoised SAR imagery.

A vegetation-type density image map was produced for the SICS area, and several 1:24,000-scale image photomap examples were printed. Classification results are promising, and further study of this technique applied to higher spatial resolution hyperspectral imagery is suggested.

(This abstract was taken from the Greater Everglades Ecosystem Restoration (GEER) Open File Report (PDF, 8.7 MB))

Note: PDF files require the Adobe Acrobat Reader to be read. Download the FREE Adobe Acrobat Reader ®.

Back to Project Homepage