The high spectral resolution of hyperspectral imaging (HSI) systems greatly enhances the capabilities of discrimination, identification and quantification of objects of different materials from remote sensing images, but they also bring challenges to the processing and analysis of HSI data. One issue is the high computation cost and the curse of dimensionality associated with the high dimensions of HSI data. A second issue is how to effectively utilize the information including spectral and spatial information embedded in HSI data.
Geometric Active Contour (GAC) is a widely used image segmentation method that utilizes the geometric information of objects within images. One category of GAC models, the region-based GAC models (RGAC), have good potential for remote sensing image processing because they use both spectral and geometry information in images are robust to initial contour placement. These models have been introduced to target extractions and classifications on remote sensing images. However, there are some restrictions on the applications of the RGAC models on remote sensing. First, the heavy involvement of iterative contour evolutions makes GAC applications time-consuming and inconvenient to use. Second, the current RGAC models must be based on a certain distance metric and the performance of RGAC classifiers are restricted by the performance of the employed distance metrics.
According to the key features of the RGAC models analyzed in this dissertation, a classification framework is developed for remote sensing image classifications using the RGAC models. This framework allows the RGAC models to be combined with conventional pixel-based classifiers to promote them to spectral-spatial classifiers and also greatly reduces the iterations of contour evolutions. An extended Chan-Vese (ECV) model is proposed that is able to incorporate the widely used distance metrics in remote sensing image processing. A new type of RGAC model, the edge-oriented RGAC model, is also discussed. The RGAC classifier utilizing this new model can be combined with any pixel-based classifier and is much more flexible and adaptive than the ECV models that must be based on a certain distance metric. Classification experiments were performed on two HSI datasets, and the proposed RGAC classifiers were evaluated and compared based on the results of the experiment.
In order to handle the curse of dimensionality of HSI, the nonlinear dimensionality reduction (DR) was tested in the research portion of this dissertation for its ability to distinguish the intrinsic nonlinear structures in HSI. An algorithm of the fast near neighbor search in high-dimensional spaces, locality-sensitive hashing (LSH), is introduced to the nonlinear DR method of Laplacian eigenmaps (LE) for speeding up the k-nearest neighbor search, which is highly computationally expensive and acts as a bottleneck for the nonlinear DR methods. Experimental results demonstrated that the LE combined with LSH perform nonlinear DR on relatively large HSI datasets, and the quality of the DR results are better than the DR results of PCA for hyperspectral image classifications.