Short Time Spatiotemporal Forecasting using Efficient Deep Learning Algorithms

Abstract

In recent times, the importance of spatiotemporal data has surged, largely due to technologies such as mobile devices, GPS systems, online mapping, weather services, and various sensor technologies. These advancements have led to significant applications in fields like Traffic and Weather forecasting.

Traffic forecasting, a crucial aspect of mobility management, has seen extensive research, both in terms of time series and spatiotemporal dynamic forecasting. Similarly, Earth Observatory research, particularly in short-term "Now-casting," has gained traction with the integration of artificial intelligence.

Traffic and weather forecasting usually involve high-resolution data, which is essential for accurately quantifying real-world distances. Also, an efficient network for these applications is needed to adequately capture interrelationships between spatial and temporal variables among others. Addressing these challenges requires developing highly efficient networks that strike a balance between parameter efficiency and computational complexity. This research proposes a unique approach by combining hypercomplex deep learning and the vision transformer architecture.

Hypercomplex numbers, such as Quaternion and Octonion, offer advantages over real numbers in convolutional neural networks. For example, Sedenion, a 16-component hypercomplex number, can significantly reduce the number of parameters while capturing complex interrelationships between its components. This feature makes hypercomplex deep learning suitable for multimodal tasks where each component corresponds to a modality.

Transformers networks, known for their self-attention mechanisms, effectively manage interrelations between token embeddings, particularly in natural language processing. This property has been harnessed in the vision transformer architecture, specially designed for visual data, where image patches are encoded into embeddings.

This dissertation explores the application of hypercomplex numbers and transformers in spatiotemporal prediction tasks, such as precise weather and traffic forecasting. Additionally, the adaptability of this approach has been assessed in segmenting marine debris, highlighting its potential in diverse applications.

Date of Award	14 Dec 2023
Original language	American English
Supervisor	Hasan Al Marzouqi (Supervisor)