Exploring Deep Learning Architectures for Spatial Interpolation

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Exploring Deep Learning Architectures for Spatial Interpolation"

By

Miss Jia LI


Abstract:

Spatial interpolation is a critical task widely studied in many fields, such as 
climate science, water science, and geography, aiming to estimate the values at 
any target location based on observations from known locations. However, 
obtaining accurate interpolation results is a nontrivial task due to the 
complexity of spatial relations between locations. The performance of 
traditional spatial interpolation techniques is limited by rigid formulations 
or subjective assumptions, which do not always hold in complex real-world 
situations. In recent years, spatiotemporal deep learning has seen impressive 
strides, however, little attention has been paid to spatial interpolation.

In this thesis, we explore deep learning architectures for spatial 
interpolation. First, we set our sights on the Graph Neural Networks (GNNs) 
architecture and propose GSI, a transductive GNN-based model that focuses on 
learning the spatial message-passing mechanism to perform effective 
interpolation. Specifically, GSI employs two major techniques: constraining the 
flow of message passing and adaptive graph structure learning. The former 
allows for effective direct interpolation when attributes are lacking, while 
the latter enables GSI to adaptively model complex spatial correlations. 
Additionally, we introduce a residual correction mechanism to refine the 
results, thereby further decreasing estimation errors. Second, we propose SSIN, 
a self-supervised learning framework to improve spatial interpolation by mining 
latent spatial patterns in historical observation data. Specifically, we 
propose the SpaFormer model as the core component of SSIN to overcome the 
limitations of pre-settings in existing methods. SpaFormer can learn 
informative embeddings for raw data, then adaptively model interactions and 
aggregate spatial context information for interpolation, instead of relying on 
any prior knowledge to characterize spatial correlations. To better investigate 
the effectiveness of our proposed solutions and other baselines, we select 
rainfall spatial interpolation as our primary case, which is a representative 
problem with a significant real-world impact. To conduct research on this task, 
we collect and process two large real-world rain gauge datasets. Finally, 
experimental results demonstrate the effectiveness of our proposed solutions 
against state-of-the-art methods.


Date:                   Thursday, 22 August 2024

Time:                   10:00am - 12:00noon

Venue:                  Room 3494
                        Lifts 25/26

Chairman:               Prof. Jeffrey Robert CHASNOV (MATH)

Committee Members:      Prof. Lei CHEN (Supervisor)
                        Prof. Qiong LUO
                        Prof. Ke YI
                        Dr. Chengxing ZHAI (EMIA)
                        Prof. Hong CHENG (CUHK)