Information content analysis and noise characterization in remote sensing image interpretation

The maximum information obtainable from an image is limited primarily by the quality of the data. The information content must be characterized to properly select the data needed and the related costs to satisfy the requirements of a given application. This dissertation examines variables such as spatial resolution, spectral resolution, and noise and how they relate to image information content. Algorithms are developed which estimate the noise standard deviation in addition to modeling and quantifying the amount of image information content. As noise is a common degradation which affects image interpretability, a novel method of noise standard deviation estimation was developed in this dissertation using data masking. The technique uses Laplacian and gradient convolution filters and compares favorably to existing methods of noise estimation. By examining the effects of noise on principal component analysis (PCA), it is observed that PCA is more effective in reducing uncorrelated normally distributed additive and multiplicative noise as compared to correlated forms of noise. This is demonstrated by noting the change in the PCA standard deviation as a function of the amount of noise. Quantification of the information content was established using an interpretability-based approach with a utility index model. The model compares the results of degraded and non-degraded imagery using a k-means unsupervised classifier. The information content is defined to be the relative classification accuracy of the imagery. The model is shown to be robust for widely varying data types and scenes. Application of the information content model to urban growth, lake surface area change, and fire burn severity verifies the effectiveness and the wide applicability of the model