Research On The Hyperspectral Image Classification Algorithms Based On Watershed Segmentation

Hyperspectral image classification is the main research topic in hyperspectral image processing and analysis, and has been widely applied to military surveillance, environmental monitoring, mineral recognition, etc. Therefore, it has received much attention from researchers. Hyperspectral image classification based only on spectral information is hard to improve accuracy, because there exist "same object with different spectra" and "same spectrum within different objects" scenarios in hyperspectral images. So, researchers have studied deeply in hyperspectral image classification using both the spectral and spatial information, and have made satisfactory progress. Besides, it is very expensive to acquire labeled hyperspectral data. How to make use of unlabeled data in hyperspectral images is worth deeply studying. Therefore, this thesis mainly focuses on mining and utilizing spatial information in hyperspectral images, and using unlabeled data in hyperspectral images. The main contributions are as follows:1. Propose an algorithm for hyperspectral image classification based on Watershed segmentation and SVM (WSVM). WSVM performs watershed segmentation on the image and applies SVM to classify the pixels at first. Then according to different segmented regions the classification results of SVM are processed using two strategies:â‘  If there exist training samples in one region and the training samples belong to the same class, all the samples in this region are labeled the same class with training samples.â‘¡ Otherwise, vote the results of SVM and label all the samples in this region to the class that the majority samples belong to. To verify the effectiveness of the newly proposed algorithm, we perform experiments on two datasets:Indian Pine and the University of Pavia (PU) datasets. Experimental results show that our algorithm is superior to some other algorithms which take into consideration the spatial information.2. Propose two algorithms for hyperspectral image classification based on watershed segmentation and sparse representation:1) the classification algorithm for hyperspectral images based on Watershed segmentation and Joint Sparse Representation (WJSR). This algorithm performs watershed segmentation on the image at first, two following classification strategies are employed according to different segmented regions:â‘  If there exist labeled samples in one region and the labels are the same, all the samples in this region are labeled the same class with the labeled samples. â‘¡ Otherwise, classify all the samples in this region using joint sparse representation.2) the classification algorithm for hyperspectral images based on Watershed segmentation and the Improved Joint Sparse Representation model (WIJSR). This algorithm performs watershed segmentation on the image at first, two following classification strategies are employed according to different segmented regions: â‘  if there exist labeled samples in one region and the labels are the same, all the samples in this region are labeled the same class as the labeled samples. â‘¡ Otherwise, WIJSR uses the Improved Joint Sparse Representation model (IJSR) to compute the coefficient of every sample in the region and then classify the samples within this region using the coefficients of all the samples, where IJSR adds a Laplace term in the existing joint sparse representation model, thus making it better sparsely represent the samples lied in the boundaries. Experimental results on Indian Pine and the University of Pavia (PU) datasets show that our algorithms have better performance than other algorithms for hyperspectral image classification based on sparse representation.3. Propose a Semi-Supervised Classification algorithm based on Watershed segmentation (WSSC). The main idea of WSSC is as follows: â‘  watershed segmentation is utilized on the image to obtain different regions. In each region, if there exists labeled samples and the labels are the same, then label all the samples in the region to the class that the labeled samples belong to; â‘¡ Construct a graph which contains spatial information of the samples via joint sparse representation based on watershed segmentation, and propagate the labels based on this graph; â‘¢ Add the samples of the same labels obtained by steps 1 and 2 to the set of labeled samples, Use the newly obtained set of labeled samples to train SVM, and classify all the unlabeled samples; â‘£ To make the classification-result-graph more smooth, we vote the classification results in each region. Finally, we perform experiments on Indian Pine and the PU datasets. The newly developed algorithm has higher classification accuracy as compared to other semi-supervised algorithms for hyperspectral image classification.