MODELING AND CLASSIFICATION OF TEXTURE IN FOREST LANDSCAPES, WITH APPLICATION TO REMOTE SENSING (MEASURES, SPATIAL INTERACTION MODELS, CHANNEL SELECTION, SPATIAL POINT PROCESSES, SIMULATION)

Choice of texture measures, channel selection, directionality, and window size are problems presented in classification of forest landscapes using texture information. The objective of this study is to provide modeling procedures for simulating image data of forest landscapes, and through such models to assess potential contribution of texture information to classification of forests from remotely sensed digital data.;The simultaneous autoregressive (SAR) model was superior to the conditional Markov model for modeling the five vegetative cover types based on quantitative image quality measures. The performances of the SAR model were acceptable.;A selected menu of texture measures was evaluated according to this procedure, and four of these measures were shown to be useful for discriminating the five vegetative cover types. These measures were: the mean norm length (MNL), the range of norm length, the variance of norm length, and the mean Euclidean distance. A good choice for classification proved to be a combination of MNL textural channels and the original spectral channels.;The negative effects of directionality can be avoided by using combined MSS channels and their MNL textural channels. If the width of every vegetative cover type was not less than twice the size of one side of a moving window, the combined MNL textural channels and spectral channels increased classification accuracy about 10%.;Five vegetative cover types were modeled using spatial interaction models based upon image segments extracted from airborne multispectral scanner (MSS) imagery. Simulations of synthetic forest landscapes were performed by spatial point processes, areal processes, and spatial interaction models. Point patterns of a natural forest were simulated by a contagion process, and inhibition process, a Poisson cluster process, and heterogeneity. The effects of texture measures, channel selection, directionality, and window size for discriminating five vegetative cover types were tested on the simulated images.;Rather than evaluating only one factor, a good strategy in practical applications of texture information is to evaluate the effects of each of the following factors: texture measures, channel selection, directionality, and suitable window sizes.