A wide range of methods for analysis of airborne- and satellite-derived imagery continues to be proposed and assessed. In this paper, we review remote sensing implementations of support vector machines (SVMs), a promising machine learning methodology. This review is timely due to the exponentially increasing number of works published in recent years. SVMs are particularly appealing in the remote sensing field due to their ability to generalize well even with limited training samples, a common limitation for remote sensing applications. However, they also suffer from parameter assignment issues that can significantly affect obtained results. A summary of empirical results is provided for various applications of over one hundred published works (as of April, 2010). It is our hope that this survey will provide guidelines for future applications of SVMs and possible areas of algorithm enhancement.

Enhanced techniques for processing and extraction of features from images, combined with robust classification algorithms can be of paramount importance in handwriting recognition. Such methods are treated in this report for automatic recognition of scanned offline cursive handwriting. Specifically, a wordbased approach to feature extraction and learning vector quantization machine learning algorithm are applied to construct a recognizer for historic handwriting from the Royal Dutch Queen’s archive. Included in this paper is a description of the results of our experiments together with an overview of the methods used for data acquisition and representations, construction and training of the classifier network, and recognition of an unseen instance.

We study, discuss and write down in this thesis report the methods and experimental results of improved and novel learning techniques aimed at providing better data class quantization by convex optimization. The concept of adaptive metrics, realized by semi-definite programming on functions parameterized by distances between data vectors has originally been shown by Weinberger (1) to improve k -nearest neighbour classifier performance. This idea has been extended to the prototype-based scenario, a more robust and intuitive iterative learning algorithm suitable for multi-way classification of data instances whose (often dauntingly overlapping) clusters are regarded as voronoi tessellations in spaces of higher dimensions. In our supervised learning scheme, placement of the reference data vectors within some defined radius in the space of training patterns of respective matching labels, and the displacement of differently labelled prototypes far away from the stimuli aim at optimizing the combined transformations with the overall goal of realizing classifiers with better limiting performance behaviour. Analogous to margin maximization learning concepts in support vector machines, the training heuristic is modelled as a constrained optimization problem involving penalty and barrier functions in the context of non- linear programming.

The availability of higher spatial, spectral and/or radiometric resolution images acquired from state- of-the-art airborne and space-borne remote sensing missions means that more detailed and potentially useful information can be derived from the imagery. These developments have spawned a whole new set of cross- disciplinary interests among researchers involved i n the field of remote sensing of the environment an d related themes. It is increasingly becoming importa nt to identify, from remote sensing data, distinct areas or features of interest such as forest, asphalt, water , grassland and buildings and assign all occurrence s of such features to distinct classes. Application domains o f the classification are numerous and include autom atic population estimation, vegetation mapping, urban he at island analysis and fighting water hyacinth, amo ng others. Modern classification techniques should the refore be robust and efficient enough to handle mul ti-band image processing. For completeness, this paper revi sits the basics of the SVM method and includes a re view of some of the recent developments in support vecto r machines suitable for membership quantization of remotely sensed data. A description and analysis of results of SVM classification on four band (blue, green, red and near-infrared) QuickBird satellite imagery of Ls Vegas acquired in May 2003 are also included and compared with the performance competing machine learning techniques.