Lightweighting the prediction process of urban states with parameter sharing and dilated operations

Lightweight and high-precision prediction models for urban states are anticipated to run efficiently on resource-limited devices, serving as key technologies for realizing smart city management. However, many existing models, despite achieving high prediction precision, suffer from overly complex de...

Full description

Saved in:
Bibliographic Details
Main Authors: Peixiao Wang, Haolong Yang, Hengcai Zhang, Shifen Cheng, Feng Lu, Zeqiang Chen
Format: Article
Language:English
Published: Taylor & Francis Group 2025-08-01
Series:International Journal of Digital Earth
Subjects:
Online Access:https://www.tandfonline.com/doi/10.1080/17538947.2025.2468414
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Lightweight and high-precision prediction models for urban states are anticipated to run efficiently on resource-limited devices, serving as key technologies for realizing smart city management. However, many existing models, despite achieving high prediction precision, suffer from overly complex designs, leading to low computational efficiency, a large number of learnable parameters, and difficulty in hyper-parameter calibration. In this study, we present a lightweight parameter-shared dilated convolutional network (PSDCN) to address these challenges. Specifically, we define parameter-shared temporal/graph dilated convolution operators to efficiently and accurately capture spatio-temporal correlations without significantly increasing model's computation time and scale of learnable parameters. Furthermore, we establish mathematical relationships between hyperparameters, significantly reducing their number and simplifying the calibration process. The PSDCN model was validated using PM2.5, traffic, and temperature datasets. The results demonstrated that the PSDCN model simplifies hyperparameter calibration. It also either outperforms or matches the prediction accuracy of nine baselines, while achieving better time efficiency and requiring fewer learnable parameters.
ISSN:1753-8947
1753-8955