Spatial frequency analysis of multispectral data.

This paper presents the definitions of texture dependent features which can be obtained in terms of the spatial frequencies of small sections of remotely sensed multispectral data. The features are made independent of the direction of view by defining them as symmetric functions of the spatial frequencies sensed with various viewing directions. Several textural features are defined and experimental results indicating existence of signatures in these features are presented. Preliminary experiments have been performed on the classification of 60 samples, 10 from each of the following 6 categories - grass, trees, water, staked tomatoes, treated ground tomatoes, and untreated ground tomatoes. Classifications of the training samples using only one feature at a time indicate that several of the features yield classification efficiencies higher than 65%. The efficiency increases considerably when combinations of these features are used.