Inductive Spatiotemporal Graph Convolutional Networks for Short-Term Quantitative Precipitation Forecasting

Yajing Wu, Xuebing Yang, Yongqiang Tang, Chenyang Zhang, Guoping Zhang, Wensheng Zhang

IEEE Transactions on Geoscience and Remote Sensing · 01 January 2022 · China Meteorological Administration (CMA)

In one sentence. Introduces InstGCN, an inductive spatiotemporal graph network for radar-based short-term rainfall forecasting, with a novel elliptic graph construction and subgraph decomposition for unseen nodes.

Abstract

Short-term quantitative precipitation forecasting (SQPF) using weather radar is an important but challenging problem as one must cope with inherent nonlinearity and spatiotemporal correlation in the data. In this article, we propose a novel deep learning model, named Inductive spatiotemporal Graph Convolutional Networks (InstGCNs), to overcome these issues in SQPF. The proposed InstGCN can learn a nonlinear mapping from historical radar reflectivity to future rainfall amounts and extract informative spatiotemporal representations simultaneously. Specifically, we first provide a formal definition for formulating the SQPF problem from a graph perspective. Then, based on radar reflectivity and rain gauge observation, we propose a novel graph construction approach that utilizes a special elliptic structure to model the spatial dependence of precipitation areas. In addition, a new Node level Differential Block (Node-DB) is introduced to tackle the nonstationary temporal dependence. To execute inductive graph learning for unseen nodes, we design to decompose a whole graph into subgraphs. We conduct extensive experiments on three real-world datasets in East China and a public weather radar dataset in the southeastern parts of France. The experimental results confirm the advantages of InstGCN compared with several state of the arts.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.