Fusing high-resolution multispectral image with trajectory for user next travel location prediction

Predicting user future travel trajectories is crucial for transportation planning, emergency evacuation, geo-customized advertising and other city smart management. However, due to the spatiotemporal complexity of human travel behavior, accurately predicting user travel trajectory is still very challenging. Existing prediction approaches majorly focus on historical trajectory pattern mining and training model variants. However, the functional land use of user stops, which may provide valuable inputs for user current and historical travel purposes, is still rarely utilized in travel prediction. On the other hand, identifying functional land use from multispectral observations has been well-developed by urban remote sensing community in recent years. Targeting this research gap, this paper presents a geospatial artificial intelligence (GeoAI) based approach for travel location prediction by fusing multispectral observations and trajectories geocomputation (FMT model). First, user stay points are extracted from raw historical trajectories with link construction, stay point extraction and semantic location clustering. Then, high-resolution multispectral images, POIs and road networks were utilized to establish a block-level functional land use identification model for every stay point, which is utilized in the trajectories prediction model with interfered travel purpose inputs. Finally, a geospatial customized neural network model with the aforementioned prediction features integrated was implemented for trajectory prediction capabilities. The prediction performance evaluations show that a) fusing multispectral images with functional land use contributed to prediction performance gains of 9.03 %, 7.14 % and 8.69 % in prediction precision, recall and F-score respectively, b) sensitivity analysis showed that the FMT model researched over 90 % of the model prediction capabilities with 40 %+ training size and 4+ time steps in our study area, c) FMT model has 12.2 % and 7.7 % performance gain as compared with traditional conventional N-Order Markov (NM) models and the Prediction by Partial Match (PPM) models.