This research paper proposes two different methods for removing biases from geo-spatial trajectories obtained from AIS, enabling the creation of machine learning models that generalize well to areas with no training data. We focus here on the task of behavior detection such as moored, underway or drifting from AIS. The first method utilizes data augmentation techniques specifically designed for augmenting geo-spatial trajectories that include angled information such as direction, while the second method is based on engineering features that eliminate these geographical biases. These two methods are compared using different types of machine learning models, including random forests, hidden Markov models, LSTMs, and transformer networks.