Abstract
The rising impact of freight-related activities from heavy commercial vehicles has recently gained the attention of urban authorities and policymakers due to their downstream negative externalities. However, past attempts at studying their operational patterns have been marred with data challenges related to its high data procurement cost, heterogeneous data storage and sharing practices between stakeholders, and the omission of critical information due to the voluntary nature of truck driver surveys. All of these factors contribute to the collection of sparse and incomplete datasets, which provide limited insights and impede research progress in the field of urban freight. With the ubiquitous adoption of mobile devices with location-aware technology and the widespread application of machine learning in various fields, this thesis’s overall objective is to apply and develop different machine learning and deep learning techniques to resolve the prevalent data challenges encountered in urban freight studies. In achieving this objective, we are also interested in extracting new and valuable insights from the data to inform decision-making and urban planning. This thesis explores three real-world studies and their respective data challenges. The first study is a commercial vehicle parking and observation survey conducted to study the delivery activities at nine urban retail malls in Singapore. An analysis of the initial dataset highlighted issues related to incomplete data, which severely limited its usefulness during post-analysis. A novel imputation and regression model was proposed to impute the incomplete fields in the dataset before predicting the commercialvehicles’ parking duration during delivery activities. Based on the model developed, we identified three significant factors related to vehicle dwell time (i.e., activity type, parking location, and volume of goods delivered) and proposed several recommendations to enhance the retail malls’ existing parking management policies. The second study investigates the data sparsity issues in existing geospatial datasets by proposing a point-of-interest (POI) conflation framework to unify the geospatial data from multiple data sources to obtain a comprehensive dataset with different urban freight applications. The third study is a commercial vehicle travel survey that utilises a digital survey platform to capture the movement patterns and parking behaviours of commercial vehicles operating within Singapore. However, due to the repetitive nature of the driver verification process, the survey suffered from a low participant response rate with a significant portion of unverified stops. Given these issues, we proposed a stop activity recognition model to pre-populate the activity-related fields during driver verification and reduce respondent burden. The model can also recover the activity information from the unverified stops to support downstream analysis.