Towards rainstorm event identification: A transfer learning framework using citizen-report texts and multi-source spatial data

Abstract

Urban rainstorms are often accompanied by various hazardous events (e.g., water inundation and small-scale debris flow) and subsequent consequential events (e.g., infrastructure failure and trapped citizens). It is imperative to achieve rapid identification of such events for timely and effective rainstorm disaster management. However, due to small-sample event classes, it becomes challenging to achieve the desired identification results. In this paper, a cross-domain transfer learning scheme is proposed to identify rainstorm events by combining citizen-report texts with multi-source spatial data. In the proposed framework, we recommend joint distribution adaptation (JDA) as the key to improving the identification of target events with inadequate samples by transferring knowledge from similar event domains. Meanwhile, multi-source spatial features are extracted and embedded in the text with linear discriminant analysis (LDA) to overcome the feature incompleteness that is caused by short text length. The proposed approach is verified based on the real data in Wuhan, China. Experimental results show that the knowledge transfer between different event classes helps cities to improve the performance of rainstorm event identification, with the accuracy increasing from 81% to 92%. Under data imbalance scenarios, the proposed scheme outperforms the state-of-the-art methods, such as sampling-based methods, ERNIE, and TCA. At the same time, the combination of textual and spatial features results in significant improvement in event identification, as evident by an 8% increase in accuracy. This framework can be followed to build a rainstorm event warning and assistance system to rapidly adapt management responses.