Multispectral remote sensing techniques are being used to characterize vegetation cover in the Florida Everglades. Information on vegetation cover, such as type, density, and mixture composition is needed at scales appropriate for large-scale hydrodynamic models to help quantify the resistance that vegetation gives to water flow (Lee, 1996).

In this application, the classification of vegetation cover from moderate-resolution sensor systems, such as LANDSAT thematic mapper (TM), is difficult because the requisite categories are mixtures of cover types at various densities; for example, sparse sawgrass and water. These mixture proportions, and thus the respective multispectral signatures, vary throughout the scene. Consequently, there may be significant overlap between the distributions of spectral signatures, such as from water and vegetation (Schowengerdt, 1983). These mixture distributions can make accurate classification difficult, for example, when using Bayes' classification methods.

Unsupervised clustering techniques can help to determine natural groupings of the spectral signa-tures that correspond to these mixture categories. Commonly, subsequent manual assignment of individual clusters or groups of clusters to particular cover classes is used to create the final vegetation classification.

In contrast, the classification technique described here is a combination of unsupervised, fuzzy clustering followed by the supervised classification of fuzzy clustered data by means of a trained multilayer perceptron neural network (NN) classifier. A major advantage is that by means of an automated (that is, the NN learning phase) instead of a manual process, the NN learns how to group the fuzzy clustered data so that classification error is minimized.

First, a fuzzy clustering technique (Bezdek, 1993) is applied to preprocessed TM multispectral data. The result, for each multispectral data sample vector, is a set of fuzzy membership values to all cluster prototype vectors. The next step is the supervised, nonparametric statistical classification of the fuzzy membership data by multilayer NN. The NN was trained during a learning phase to develop a mapping between the fuzzy clustered data and the desired output class by means of teaching/training pair examples.

NN training data consist of pairs of input data samples and the desired class. Vegetation class type and density ground truth information at the 30-m spatial resolution of the TM sensor were inferred by aggregating very high spatial resolution (0.5 m) vegetation class samples. The class information was derived from data collected by an airborne, digital multispectral video system (Anderson and others, 1997). Multispectral and derived feature data used as input to the classification process are TM band 3 (tm3), normalized difference vegetation index (tm4-tm3)/(tm4+tm3), and spatial resolution-enhanced thermal-infrared (T-IR) data, tm6 (Lemeshewsky, 1997).

This classification process is similar to the supervised merging (consistent with ground truth) of hard clustered data into classes with, however these distinctions: the process of grouping, or merging, fuzzy (instead of hard) clustered membership feature data by means of the NN is such that the classification error is minimized, and this process is automatically learned during the NN training process.

Six TM multispectral data bands have 30-by-30-m instantaneous field of view (IFOV), but the T-IR band has 120- by- 120-m IFOV. In this study, an NN-based technique was developed to enhance the spatial resolution of the T-IR data to 30 m. In a multistep process, a multilayer NN was trained so that its output approximates T-IR data at 120-m sample distance as a function of TM bands 7, 5, and 4 input data that have been resampled (reduced in resolution) to 120-m sample distance. The output of this trained NN, for 30-m resolution input, gives an approximation of T-IR data at a higher resolution; that is, 30 m. This output was then used to enhance the resolution of the raw T-IR data to 30 m.

Preliminary results of vegetation classification tests were obtained and benefits of using the improved resolution T-IR data were evaluated. Benefits of this classification technique include the automated (and optimal in terms of minimal mean square error) mapping of clustered data to vegetation density classes and the potential for improving results by integrating other data into the classifier by means of NN training.

REFERENCES

Anderson, J.E., Desmond, G.B., Lemeshewsky, G.P., and Morgan, D.R., 1997: Reflectance calibrated digital multispectral video: A test bed for high spectral and spatial resolution remote sensing: PE&RS, v. 53, no. 3, p. 224-229.

Bezdek, J.C., 1993: A review of probabilistic, fuzzy and neural models for pattern recognition: Journal of Intelligent and Fuzzy Systems, v. 1, no. 1, p.1-25.

Lee, J.K., 1996: Vegetation affects water movement in the Florida Everglades: U.S. Geological Survey Fact Sheet FS-147-96.

Lemeshewsky, G., 1997: Neural network method for sharpening LANDSAT thermal data from higher resolution multispectral data: U.S. Geological Survey Open-File Report 97-301.

Schowengerdt, R.A., 1983: Digital image classification, chapter 3 of Techniques for image processing and classification in remote sensing: Academic Press, Orlando, Florida, p. 129-207.

Back to Project Homepage