Reliability of unattended ground sensors (UGS) to detect and classify different activities (e.g., walking and digging) is often limited by high false alarm rates, possibly due to the lack of robustness of the underlying algorithms in different environmental conditions (e.g., soil types and moisture contents for seismic sensors), inability to model large variations in the signature of a single activity and limitations of on-board computation. In this regard, a fast and robust multi-scale symbolic time series analysis (MSTSA) framework has been formulated to detect and classify human activities from seismic signatures. The building block of the proposed framework is built upon the concept of applying the short-length symbolic time-series online classifier (SSTOC) via Dirichlet-Compound-Multinomial model (DCM) construction. The algorithm operates on symbol sequences that are generated from seismic time-series and intermediate event class time-series at different time scales. These building blocks, with different window sizes, are cascaded in multiple layers for event detection and activity classification. A variety of experiments have been conducted in the field, which include realistic scenarios of different types of walking/digging. The results of experiments show that an accuracy of more than 90% and a false alarm of around 5% can be achieved in real time for activity detection and recognition.