High-Accuracy And Parallel Algorithms For Supervised Remote-sensing Image Classification

As a synthetic subject, remote sensing has gotten rapid development in recent decades. It plays an important role in civil and military applications as it provides a unique perspective from which to observe large regions and acquire abundant useful information in high speed. Remote-sensing image processing is dedicated to processing and analyzing digital images acquired by remote sensors. As a primary quantitative means for image analysis, supervised remote-sensing image classification is a crucial issue during the whole procedure of image processing, and has comprehensive applications in many domains.With the development of remote sensors, remote platforms, and data communication technologies, more and more remote sensing data are produced for processing. At the same time, applications have an increasing demand for high processing speed and accuracy. Thus, fast processing technologies for supervised remote-sensing image classification with high accuracy become more and more urgent, and come to be the focus of research.The procedure of supervised remote-sensing image classification can be divided into three phases: preprocessing phase, learning and discrimination phase, postprocessing phase. The learning algorithm of the second phase is critical for the accuracy of classification. Fastening geometric correction in the first phase and supervised learning step in the second phase are two key issues to improve processing speed. This thesis aims at improving the classifying accuracy and processing speed for supervised remote-sensing image classification, and makes the following contributions:1. A general coding method SCM is presented, which can generate error-correcting output codes suitable for the classification problem with any number of classes.Reducing a multi-class problem to multiple two-class problems can extend the applications for many learning algorithms. Error-correcting output codes (ECOCs) approach has this function as well as improve classifying accuracy. However, there is no coding method that can generate ECOCs suitable for any number of classes. Thus, this thesis proposes a search coding method (SCM) that associates nonnegative integers with binary strings. Given any number of classes and an expected minimum hamming distance, SCM can find out a satisfying output code through searching an integer range. Applied to several supervised learning algorithms, SCM is demonstrated to improve the recognition accuracy for both stable and unstable classifiers efficiently.2. A structured neural network based on SCM (CSNN) is explored.Serpico and Roli have proposed two neural networks with special structures devoted tomulti-sensor image recognition. This thesis extends the neural networks to a structured neural network (SNN) and a combined structured neural network (k-SNN). SNN has deterministic network structure and interpretable network behavior. However, its classifying accuracy is not high enough. Though K-SNN can improve the recognition accuracy of SNN to some extent, its learning time increases significantly. Based on the search coding method SCM, a structured neural network (CSNN) is presented; meanwhile a simple method to interpret the behavior of the structured neural networks is proposed. Experimental results show that CSNN gets the best classifying accuracy among the three structured neural networks while retaining the intelligibility.3. A dynamic range-splitting method based on continuous attributes transform (RCAT) is explored. Based on RCAT, a binary-classification tree system (Btrees) is implemented.Gray values of multi-spectral images are main attributes for supervised remote-sensing image classification. Discretization of these continuous attributes has a great impact on the classification result for many supervised learning algorithms. This thesis explores a dynamic discretization method named RCAT. RCAT uses simple binarization to get a multi-splitting result through mapping a continuous attribute into a probability attribute, where the binarization of the probability attribute corresponds to a mu...