An Effective Approach to Predicting Large Dataset in Spatial Data Mining Area

Due to enormous quantities of spatial satellite images, telecommunication images, health related tools etc., it is often impractical for users to have detailed and thorough examination of spatial data (S). Large dataset is very common and pervasive in a number of application areas. Discovering or predicting patterns from these datasets is very vital. This research focused on developing new methods, models and techniques for accomplishing advanced spatial data mining (ASDM) tasks. The algorithms were designed to challenge state-of-the-art data technologies and they are tested with randomly generated and actual real-world data. Two main approaches were adopted to achieve the objectives (1) identifying the actual data types (DTs), data structures and spatial content of a given dataset (to make our model versatile and robust) and (2) integrating these data types into an appropriate database management system (DBMS) framework, for easy management and manipulation. These two approaches helped to discover the general and varying types of patterns that exist within any given dataset non-spatial, spatial or even temporal (because spatial data are always influenced by temporal agents) datasets. An iterative method was adopted for system development methodology in this study. The method was adopted as a strategy to combat the irregularity that often exists within spatial datasets. In the course of this study, some of the challenges we encountered which also doubled as current challenges facing spatial data mining includes: (a) time complexity in availing useful data for analysis, (b) time complexity in loading data to storage and (c) difficulties in discovering spatial, non-spatial and temporal correlations between different data objects. However, despite the above challenges, there are some opportunities that spatial data can benefit from including: Cloud computing, Spark technology, Parallelisation, and Bulk-loading methods.

Techniques and application areas of spatial data mining (SDM) were identified and their strength and limitations were equally documented. Finally, new methods and algorithms for mining very large data of spatial/non-spatial bias were created. The proposed models/systems are documented in the sections as follows: (a) Development of a new technique for parallel indexing of large dataset (PaX-DBSCAN), (b) Development of new techniques for clustering (X-DBSCAN) in a learning process, (c) Development of a new technique for detecting human skin in an image, (d) Development of a new technique for finding face in an image, (e) Development of a novel technique for management of large spatial and non-spatial datasets (aX-tree).

The most prominent among our methods is the new structure used in (c) above -- packed maintained k-dimensional tree (Pmkd-tree), for fast spatial indexing and querying. The structure is a combination system that combines all the proposed algorithms to produce one solid, standard, useful and quality system. The intention of the new final algorithm (system) is to combine the entire initial proposed algorithms to come up with one strong generic effective tool for predicting large dataset SDM area, which it is capable of finding patterns that exist among spatial or non-spatial objects in a DBMS. In addition to Pmkd-tree, we also implemented a novel spatial structure, packed quad-tree (Pquad-Tree), to balance and speed up the performance of the regular quad-tree. Our systems so far have shown a manifestation of efficiency in terms of performance, storage and speed. The final Systems (Pmkd-tree and Pquad-Tree) are generic systems that are flexible, robust, light and stable. They are explicit spatial models for analysing any given problem and for predicting objects as spatially distributed events, using basic SDM algorithms. They can be applied to pattern matching, image processing, computer vision, bioinformatics, information retrieval, machine learning (classification and clustering) and many other computational tasks.