Activity-based models consider travelers' behaviours as a sequence of trips and can be used to simulate travel behaviour and extract aggregate demand metrics. The current paper focuses on predicting individual trip activities and transport modes by exploring two different long short-term memory (LSTM) model architectures, and aims to augment the feature space with geospatial information using an Auto- Encoder to improve predictive performance. The geospatial information used includes descriptive spatial and mobility characteristics for each trip origin/destination location. The results show that including geospatial representations in the models' feature space enhanced the models' performance by 1% to 4% and by 4% to 5% in terms of F1 score for predicting trip activities and modes respectively compared. The suggested models can be useful for transport planners and policy-makers in making informed decisions about transport networks based on travel demand.