| Spatial-temporal data mining techniques have become increasingly important in emerging fields such as remote sensing, precision agriculture, geoscience and brain imaging. In this Thesis, novel spatial-temporal data mining methods and algorithms are presented. After the introductory remarks, modeling spatial-temporal attributes with short observation history using spatial-temporal autoregressive models on uniform grid is explored. Model specifications (including covariance structure and stationarity) are discussed as well as issues in model identification, estimation and forecasting on three different sampling schedules. The proposed technique is experimentally evaluated on simulated spatial-temporal processes that confirm to model assumptions as well as on real-life agricultural data. Subsequently, we proceed with spatial-temporal prediction of a response variable with a partial observability of influential attributes. After mathematical definition of the proposed model, evaluation of the estimation technique on synthetic data that conform to the modeling assumptions is performed and a model is assessed on simulated realistic spatial-temporal data, obtained using the proposed data generator. The following part of the Thesis is dedicated to spatial-temporal profit optimization using neural network modeling. Profit optimization is proposed using a two-phase process that consists of estimation of response/attribute dependence and profit optimization for a particular tuple of attribute values. The proposed method is evaluated on simulated precision agriculture data. Next, we introduce a spatial-temporal data simulator, which is an important tool for evaluation of knowledge discovery methods for spatial-temporal domains. Various aspects of the proposed data generator are discussed, including generation of features and simulation of response variable as well as a practical implementation of the proposed method and its application on experiments with simulated data. The conclusive part of the Thesis contains an overview of results and contributions and directions for future work. |
