Efficient Nonlinear Dimensionality Reduction for Pixel-wise Classification of Hyperspectral Imagery

Classification, target detection, and compression are all important tasks in analyzing hyperspectral imagery (HSI). Because of the high dimensionality of HSI, it is often useful to identify low-dimensional representations of HSI data that can be used to make analysis tasks tractable. Traditional linear dimensionality reduction (DR) methods are not adequate due to the nonlinear distribution of HSI data. Many nonlinear DR methods, which are successful in the general data processing domain, such as Local Linear Embedding (LLE) [1], Isometric Feature Mapping (ISOMAP) [2] and Kernel Principal Components Analysis (KPCA) [3], run very slowly and require large amounts of memory when applied to HSI. For example, applying KPCA to the 512×217 pixel, 204-band Salinas image using a modern desktop computer (AMD FX-6300 Six-Core Processor, 32 GB memory) requires more than 5 days of computing time and 28GB memory! In this thesis, we propose two different algorithms for significantly improving the computational efficiency of nonlinear DR without adversely affecting the performance of classification task: Simple Linear Iterative Clustering (SLIC) superpixels and semi-supervised deep autoencoder networks (SSDAN). SLIC is a very popular algorithm developed for computing superpixels in RGB images that can easily be extended to HSI. Each superpixel includes hundreds or thousands of pixels based on spatial and spectral similarities and is represented by the mean spectrum and spatial position of all of its component pixels. Since the number of superpixels is much smaller than the number of pixels in the image, they can be used as input for nonlinearDR, which significantly reduces the required computation time and memory versus providing all of the original pixels as input. After nonlinear DR is performed using superpixels as input, an interpolation step can be used to obtain the embedding of each original image pixel in the low dimensional space. To illustrate the power of using superpixels in an HSI classification pipeline,we conduct experiments on three widely used and publicly available hyperspectral images: Indian Pines, Salinas and Pavia. The experimental results for all three images demonstrate that for moderately sized superpixels, the overall accuracy of classification using superpixel-based nonlinear DR matches and sometimes exceeds the overall accuracy of classification using pixel-based nonlinear DR, with a computational speed that is two-three orders of magnitude faster. Even though superpixel-based nonlinear DR shows promise for HSI classification, it does have disadvantages. First, it is costly to perform out-of-sample extensions. Second, it does not generalize to handle other types of data that might not have spatial information. Third, the original input pixels cannot approximately be recovered, as is possible in many DR algorithms.In order to overcome these difficulties, a new autoencoder network - SSDAN is proposed. It is a fully-connected semi-supervised autoencoder network that performs nonlinear DR in a manner that enables class information to be integrated. Features learned from SSDAN will be similar to those computed via traditional nonlinear DR, and features from the same class will be close to each other. Once the network is trained well with training data, test data can be easily mapped to the low dimensional embedding. Any kind of data can be used to train a SSDAN,and the decoder portion of the SSDAN can easily recover the initial input with reasonable loss.Experimental results on pixel-based classification in the Indian Pines, Salinas and Pavia images show that SSDANs can approximate the overall accuracy of nonlinear DR while significantly improving computational efficiency. We also show that transfer learning can be use to finetune features of a trained SSDAN for a new HSI dataset. Finally, experimental results on HSI compression show a trade-off between Overall Accuracy (OA) of extracted features and PeakSignal to Noise Ratio (PSNR) of the reconstructed image