Adaptive parametric estimation and classification of remotely sensed imagery using a pyramid structure

The earth resources observation satellite acquires imagery at high spatial temporal resolution, thereby providing a huge number of data points to be analyzed. With a current interest in problems associated with global change, it is necessary to analyze data from large regions over a long time history. Therefore, a reduction in the number of operations performed on each high resolution image is required.; If the objects or regions of interest are relatively large compared to the pixel size, then the probability of groups of contiguous pixels being in the same class is much greater than the probability being in different classes. For large homogeneous interior regions, it may not be necessary to classify all the individual pixels to identify the class label of that region. For boundary regions between two or more classes, however, more pixels are needed to identify the exact boundary. By identifying homogeneous interior regions without searching all the pixels, the number of operation and computational cost can be reduced. One means of efficiently implementing this idea involves a multilevel approach which can be applied naturally through a pyramid structure.; This algorithm adapts the concept of unsupervised region based image segmentation using the pyramid structure. However, unlike the traditional region based image segmentations, which depend on the local merging and splitting of regions based on the similarity of neighboring regions, this algorithm identifies the homogeneous and boundary regions at each level of pyramid using region statistics. The global parameters of each class are estimated based on the values of the homogeneous regions represented at that level of the pyramid using mixture distribution estimation. These estimated parameters are implicitly used as a merging criterion of homogeneous regions. The new homogeneous regions at a given level are labeled and projected to the next lower level.; This algorithm has several advantages: (1) reduction of computational cost, (2) no ordering problems in region merging associated with global estimation and labeling, (3) a gain in efficiency by starting from very homogeneous regions and adapting the variability or change of each class along the boundary regions through the multilevel structure, (4) easy adoption to the substantial changes in characteristics of class with time and between study areas, (5) reduction of sensitivity of noise.