| Recent progress on computer and sensor technology has generated huge amounts of spatial and temporal data in the Earth Science domain including climate observations, land cover time series records, ground level pollution measurements, etc. Combined with historical climate records and predictions from ecosystem models, it offers new opportunities for understanding how the earth is changing, for determining what factors cause these changes and for forecasting future changes. Spatial and temporal data mining provides innovative solutions for mining Earth Science data by incorporating spatial and temporal dependencies into standard data mining techniques. Although there has been substantial research in spatial and temporal data mining, there are still many technical issues that need to be addressed. This includes issues such as processing massive high resolution data, reducing the effect of noise, fusing data from heterogeneous sources, etc. We develop a class of efficient and robust spatial and temporal data mining algorithms in this thesis to overcome these challenges. First, we develop an integrated localized prediction framework based on Support Vector Machine to incorporate spatial and temporal dependencies. Efficient algorithms are also proposed to reduce its computational overhead. Second, we study the error accumulation problem in multi-step time series prediction and develop a novel semi-supervised multivariate time series prediction algorithm for long term forecasting. A covariance alignment method is also proposed to reduce the inconsistencies between historical and future climate data when applying the algorithm to future climate projection problems. Third, we propose a graph-based framework to detect and categorize different types of anomalies in multivariate time series data. We applied the framework to the problem of detecting and characterizing ecosystem disturbances in Earth Science data. While the spatial and temporal data mining techniques proposed in this thesis have been applied to many problems in the Earth Science domain, they are also applicable to spatial and temporal data in other application domains such as traffic analysis, image processing, network monitoring, etc. |
