With the advances in wireless and location-acquisition communication technologies, spatial-temporal data is ubiquitous in real world ranging from social media to urban planning. Two important tasks in spatial-temporal analysis are (i) inference, e.g., estimating the data for unknown locations by taking advantage of the observations from known locations; (ii) forecasting, e.g., with the aim of predicting future trends by understanding past observations with spatial-temporal information.A key challenge in mining spatial-temporal data often lies in the complex dependence structures from spatial-temporal dimensions. To fully harness the power of spatial-temporal data, this work aims to develop novel machine learning frameworks to make inferences and predictions on data by uncovering the dynamic spatial-temporal patterns. Work in this thesis investigates various applications that help data-driven decision makers by providing a better understanding of our physical environments. The results of the work in this proposal are important because they provide a solid analytical foundation to accurate and effective modeling of spatial-temporal data, and directly contribute to the emerging field of computational sustainability, social science and urban planning.