With the rapid development of various geospatial technologies including remote sensing, mobile devices, and Global Position System (GPS), spatio-temporal data are abundantly available nowadays. Extracting valuable knowledge from spatio-temporal data is of crucial importance for many real-world applications such as intelligent transportation, social services, and intelligent distribution. With the fast increase of the amount and resolution of spatio-temporal data, traditional data mining methods are becoming obsolete. In recent years, deep learning models such as Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) have made promising achievements in many fields based on the strong ability in automated feature extraction and have been broadly used in different spatio-temporal data mining tasks. Many methods have been developed, and more diverse data were collected in recent decades, however, the existing methods have faced challenges from multi-source geospatial data. This thesis investigates four efficient techniques in different scenarios for spatio-temporal data mining that take advantage of multi-source geospatial data to overcome the limitations of traditional data mining methods. This study investigates spatio-temporal data mining from four different perspectives. Firstly, a multi-elemental geolocation inference method is proposed to predict the location of tweets without geo-tags. Secondly, an optimization model is proposed to detect multiple Areas-of-Interest (AOIs) simultaneously and solve the multi-AOIs detection problem. Thirdly, a multi-task Res-U-Net model with attention mechanism is developed for the extraction of the building roofs and the whole building shapes from remote sensing images, then an offset vector method is used to detect the footprints of the high-rise buildings based on the boundaries of the corresponding building roofs and shapes. Lastly, a novel decoder fusion model is introduced to extract interior road network from remote sensing images and GPS trajectory data. And this method is effective for multi-source data mining. The proposed four methods use different techniques for spatio-temporal data mining to improve the detection performance. Numerous experiments show that the techniques developed in this thesis can detect ground features efficiently and effectively and overcome the limitations of conventional algorithms. The studies demonstrate that exploiting spatial information from multi-source geospatial data can improve the detection accuracy in comparison with single-source geospatial data.