The detection of unusual behavior plays a crucial role in the prevention of illegal and harmful activities such as smuggling, piracy, arms trading, human trafficking and illegal immigration. Also for military applications, it is useful to detect anomalous behavior to provide an alert for potential threats, especially with the more recent widespread use of drones for warfare and terrorist activities. In order to provide a solution for these emerging needs, in this work, a novel method for target dynamic behavior classification by analyzing trajectories using data gathered from multiple heterogeneous sensors is presented. The aim is to develop a context-free (i.e., sensor indifferent) method to robustly classify a selected set of anomalous trajectories and present those results to a human operator in remote sensing applications. In order to track maneuvering targets, it is typically required to use multiple (dynamic) model approach to make dynamic state inferences accurately. In practice, system parameters are typically tuned manually. The approach taken here is to estimate the parameters of a set of these so-called hybrid systems, also known as jump-Markov systems (JMS). This task will be carried out using a maximum likelihood (ML) approach, via the expectation-maximization (EM) algorithm. To handle highly nonlinear measurement models, such as those typically present in remote sensing applications, sequential Monte Carlo (SMC) or particle filters will be used to make accurate state estimations. These methods are also capable of dealing with non-Gaussian measurement or process noise. The central question explored in this work is if jump-Markov systems are a suitable class of models for trajectory classification by learning their parameters from multi-sensor measurement data.