DATA MINING TECHNIQUES ON VOLCANO MONITORING

The aim of this thesis is the study of data mining process able to discover implicit information from huge amount of data. In particular, indexing of datasets is studied to speed the efficiency of search algorithm. All of the presented techniques are applied in geophysical research field where the huge amount of data hide implicit information related to volcanic processes and their evolution over time. Data mining techniques, reported in details in the next chapters, are implemented with the aim of recurrent patterns analysis from heterogeneous data. This thesis is organized as follows. Chapter 1 introduces the problem of searching in a metric space, showing the key applications (from text retrieval to computational biology and so on) and the basic concepts (e.g. metric distance function). The current solutions, together with a model for standardization, are presented in Chapter 2. A novel indexing structure, the K-Pole Tree, that uses a dynamic number of pivots to partition a metric space, is presented in Chapter 3, after a taxonomy of the state-of-the-art indexing algorithm. Experimental effectiveness of K-Pole Tree is compared to other efficient algorithms in Chapter 4, where proximity queries results are showed. In Chapter 5 a basic review of pattern recognition techniques is reported. In particular, DBSCAN Algorithm and SVM (Support Vector Machines) are discussed. Finally, Chapter 6 shows some geophysical applications where data mining techniques are applied for volcano data analysis and surveillance purpose. In particular, an application for clustering infrasound signals and another to index an thermal image database are presented.