| Land cover characterization is one of the major applications of remote sensing imagery and it is critical for land resource management, climate modeling, and retrieval of other land cover parameters. A central issue in land cover characterization using remote sensing imagery is that of "scale." For example, land cover patches in a scene typically exist at multiple spatial scales and can consist of varying degrees of mixing of land cover classes. The main objective of this dissertation is to explore two related issues of scale that arise in the characterization of land cover from remote sensing imagery of a single spatial resolution: (i) modeling of sub-pixel land cover mixing, and (ii) modeling of supra-pixel (i.e., above pixel scale) land cover structure and mixing at both multiple spatial resolutions and multiple categorical granularities.; Gaussian mixture discriminant analysis (MDA) accounts for within-class spectral variability at the pixel scale by modeling a class distribution as a weighted sum of a few Gaussian components and it provides a means for modeling sub-pixel land cover mixing. In particular, when applied to a forest dataset from California, the MDA method performs better than a conventional linear mixture model and achieves performance similar to that of an ARTMAP neural network, while retaining its easy interpretability.; Statistical mixture models, recursive dyadic partitions, and categorical hierarchies are used to characterize relevant aspects of scale in the supra-pixel problem. Specifically, three topics on aggregation of increasing sophistication are addressed: the use of original spectral measurements across multiple spatial scales, adaptive choice of spatial scale, and adaptive choice of categorical granularity jointly with spatial scale. Evaluation with simulated imagery shows that the approaches examined provide increasing accuracy in land cover characterization and are superior to commonly used label aggregation approaches.; The multiscale multigranular framework derived above is validated using data collected for this purpose from Plymouth County, Massachusetts. Expert evaluation using fuzzy sets shows that this multiscale multigranular land cover characterization reduces thematic errors in land cover maps compared to the use of monogranular class labels. |
