Predicting freight mode choice using machine learning

In the Netherlands, several policies have been put in place to encourage a shift from freight road transportation to the less carbon-intensive alternatives of rail and inland waterways. A modal split model is needed to assess the potential impact of such policy interventions on freight mode choice. These models typically employ a Multinomial Logit (MNL) model which is straightforward to develop, transparent, and easily interpretable but suffers from low prediction accuracy if the initial assumptions made are incorrect. Currently, the potential of machine learning for more accurate freight mode choice predictions is still being explored. In my thesis, I will develop several machine learning models to predict freight transportation mode choice. I will compare and discuss the predictive accuracy and interpretability of the various models to determine whether machine learning could serve as an alternative or complement to the MNL model in NEAC, a European freight transport model. I will also explore whether other factors that affect freight decision-making could be included in the models using external data sources to further enhance the models’ accuracy. More accurate modal split models could lead to better policy recommendations, supporting both urban and regional sustainability goals by reducing road congestion and carbon emissions.

Prediction of urban noise levels: A machine learning approach combining street view images with field sound sampling

Urban noise pollution is an increasing environmental problem with significant impacts on public health and quality of life. The conventional methods of noise monitoring could not provide an adequate appraisal of the urban noise because they are relatively expensive and have sparse spatial coverage. To address this gap, I will develop a new approach comparing SVR, RF, GBRT, CNN, combining street view imagery with real-time sound sampling to predict noise levels in Dutch cities. I'll use an Insta360 X4 camera to coalesce audio and visual data from diverse spaces at three different times of the day for temporal differences with the pair of noise levels throughout the day. Visual features such as road structure, building density, and vegetation will be extracted using deep learning models (namely, Faster R-CNN, PSPNet, and CNN); acoustic metrics (namely, equivalent sound level (Leq) and Zwick loudness) will quantify noise intensity. To achieve my goal of estimating noise levels across multiple domains, I will combine these multimodal data sources and develop predictive models that apply to unmonitored areas based solely on Google Street View imagery. It will create noise maps which will enable municipalities to identify hotspots and implement targeted mitigation strategies that will make urban environments healthier.

Balancing Trade-offs between Green Spaces and Parking Accessibility in Residential Location Choice

With exponential growth of cities, the demand for suitable residential development for inhabitants is a challenge. Moreover, residential location preferences play a pivotal role in the discussion on high density urban planning. But provision of all amenities is a difficult task considering the paucity of land as a resource; therefore, residents are required to choose among them. Having an affinity to personal green spaces as well as a proximal parking space, residents certainly exhibit different priorities, having cascading implications on the spatial development paradigms. But what preferences and trade-offs do different people make within these microscopic residential choice attributes? To delve deeper into people’s preference for such accessibility and the heterogeneity thereof across segments of the population, I will conduct a stated choice experiment, in which people will experience trade-offs between residential amenities, accessibility and price (incorporating street imagery). Subsequently, I will extract insights from the data using traditional discrete choice models and state-of-the-art computer vision-enriched discrete choice models. These insights, reinforced with segment-specific preferences and sensitivities, shall be used by the municipalities to propose suitable policy interventions.

Towards sustainable urban environments: Adapting the Audio Spectrogram Transformer for soundscape analysis

My thesis aims to bridge a critical gap in sustainable urban planning by integrating advanced soundscape analysis with deep learning techniques. It focuses on adapting the Audio Spectrogram Transformer (AST), a cutting-edge, attention-based model, to predict eight perceptual attributes of urban soundscapes: Pleasant, Vibrant, Eventful, Chaotic, Annoying, Monotonous, Uneventful, and Calm. I will use the data provided by the International Soundscape Database, which includes audio recordings accompanied by in situ perceptual ratings, the research begins by converting raw audio into Mel-spectrograms, the preferred input format for the AST. The model is then adapted by modifying its final layer to output predictions corresponding to each of the eight attributes. By employing deep learning and explainable AI techniques, the research aims to provide urban planners with actionable insights to design healthier, more liveable urban environments.

Understanding Dutch citizen's preferences: a stated choice experiment on the location and design of transformer houses

The energy transition and increasing electrification require more transformer houses to ensure a stable electricity grid. Expanding their number will help meet rising energy demand (e.g., from electric vehicles) and increasing local supply (e.g., solar power). The municipality of Rotterdam has developed a framework for selecting transformer house locations, incorporating technical and some social considerations. However, the Dutch citizens’ preferences and their quantifcations regarding the visual and locational aspects of these structures are still unknown. As a result, public dissatisfaction can lead to complaints, delaying the energy transition. My thesis examines residents' preferences for both location attributes—such as placement on a parking lot versus replacing a tree—and transformer attributes like color. Through a stated choice experiment, I will analyze how in general Dutch citizens value different factors and the trade-offs they are willing to make. These insights could improve the site selection framework by incorporating social considerations or strengthening public communication on location choices.

Validating the Residential Location Choice Model: Assessing Its Alignment with Revealed Preference and previous research

My Thesis is about validating if the utility derived from the CV-model, created by Sander, does correlate with life satisfaction in the same way previous research has measured this correlation. Why is this needed: it is hard to objectively measure the physical appearence of a built environment due to the subjective nature of how people perceive things. With this CV-model this is overcome since it derives utilty from an image, making it a unique tool to quantify the quality of one's built environment, which was previously not possible. However before it can be used as a tool, it should be validated if it indeed measures the correct things and to assure it is in line with previous research.

Understanding cycling route choice behaviour through street-level images and computer vision enriched discrete choice models

It is well-known that cycling infrastructure's safety plays an important role in peoples' decisions to make a trip by bike or another mode (e.g. car). As such, safe cycling infrastructure is crucial for promoting sustainable transport. But what does safe cycling infrastructure look like, and to whom? What is considered a safe cycling infrastructure by, say, young people may feel unsafe to older people, or vice versa. To shed light on people's preferences for cycling infrastructure and the heterogeneity thereof across segments of the population, I will conduct a stated choice experiment in which people face trade-offs between travel time and cycling infrastructure (as presented using street-view images). Then, I'll analyse these data using traditional discrete choice models and the recently proposed computer vision-enriched discrete choice models. Finally, I'll use the estimated discrete choice models to produce maps showing utility-based safety scores for cycling routes in Rotterdam. I hope my study delivers insights that the municipality of Rotterdam can use to devise policies aiming to promote cycling.

Access the finished thesis via the TU Delft Repository.

Unveiling preference-based liveability in Rotterdam, exploring XAI techniques for computer vision-enriched discrete choice models

Keeping cities liveable is increasingly a challenge with increasing urbanisation. Most literature on liveability concerns perceived liveability: they study what factors, e.g. physical, social, or economic, impact peoples' liveability perceptions. Perceptions are subjective interpretations of sensory stimuli, which may influence but do not necessarily determine individuals' choice behaviour. In contrast, I study preference-based liveability. Preferences govern what people choose and do. More specifically, I aim to provide explanations for predictions made by a recently developed computer vision-enriched discrete choice model. This model produces preference-based liveability scores, taking street-view images as inputs. To achieve this, I will assess the efficacy of eXplainable Artificial Intelligence (XAI) techniques, such as LIME and Shapley. Also, I will investigate how the combination of computer vision-enriched discrete choice models and XAI can effectively be used by municipalities to support policy-making.

Access the finished thesis via the TU Delft Repository.

Investigating the impacts of built-environment features on residential location choice behaviour using computer vision techniques

Visual stimuli (i.e. images) play a crucial role in many multi-attribute decision situations, such as residential location choice behaviour. On housing platforms like Funda, street-view images showing the surrounding Built-Environment (BE) near residences under consideration offer the most direct means for individuals to understand their potential living environments. However, BE features are hard to quantify and incorporate in traditional discrete choice models as most are amorphous (i.e. without a clearly defined shape or form). This thesis proposes quantifying BE features in images into pixels or instances using panoptic image segmentation models. These pixel and instance counts, in turn, are used in a traditional discrete choice model to explain choices collected in a recently conducted residential location stated-choice experiment. Studying the extent to which visual features of the BE affect residential location choices and how people make trade-offs between BE features and numeric attributes, like cost and travel time, will offer valuable insights to urban planners.

Access the finished thesis via the TU Delft Repository.

Operationalising liveability using urban embeddings from a transport policy perspective

Transport policy is increasingly studied through the lens of broad prosperity and liveability. The Leefbaarometer offers an operationalisation of liveability, combining indicators that describe geographic areas and peoples' subjective valuations (perceived liveability). However, creating, updating, and collecting the data behind these indicators and valuations is labour-intensive. Machine learning could streamline this process and reduce the effort involved. Urban representation learning transforms complex data into simpler numerical forms for analysis. My thesis explores the creation of urban representations using diverse data sources, including street-view images and road network characteristics. It also considers how the connectivity of multi-modal transport networks (walking, cycling, driving, and public transport) can guide the learning process. These representations will be examined to determine their ability to predict Leefbaarometer scores and their components accurately.

Access the finished thesis via the TU Delft Repository.

Urban fragmentation and spatial segregation patterns in Western-Europe - A similarity analysis and identification of general trends

Transportation infrastructures play a paradoxical role in the urban space. While they facilitate the movement between certain points, they also generate fragmentation and separation. Highways, railways and congested roads, to name a few, produce barrier effects that limit the movement and interaction of people at the local scale. A certain degree of urban fragmentation is inevitable, but there is a question to be asked about its severity and the groups impacted by the resulting severance. My thesis project aims to explore the relationship between urban fragmentation patterns and residential segregation patterns of people with non-EU backgrounds across major cities in Western Europe. Spatial and statistical analysis are at the core of this research, with the purpose of identifying general trends and expanding our knowledge on the role of transportation infrastructures in segregation.

Access the finished thesis via the TU Delft Repository.