Spatial and Temporal Pattern of Rainstorms Based on Manifold Learning Algorithm

Yuanyuan Liu, Yesen Liu, Hancheng Ren, Longgang Du, Shu Liu, Li Zhang, Caiyuan Wang, Qiang GaoWater 2022 · DOI

TL;DR

This paper introduces a novel application of manifold learning, specifically Locally Linear Embedding (LLE), to analyze and extract the complex spatial-temporal patterns of rainstorms. By reducing high-dimensional rainfall data into a lower-dimensional space, the method reveals distinct rainstorm characteristics in Beijing and Shenzhen, correlating them with local topography and water vapor sources, respectively. This approach offers a more quantitative way to understand and potentially predict rainstorm behavior for improved flood forecasting and disaster prevention.

What Problem Does This Paper Address?

Urban floods, landslides, and other disasters are often triggered by heavy rainfall events. Accurately predicting these events and understanding their dynamics is crucial for effective disaster prevention. While previous research has analyzed rainfall patterns, static indicators like total rainfall amounts don't fully capture the dynamic, evolving nature of rainstorms in both space and time. Understanding the movement, intensity changes, and spatial extent of rainstorm centers is vital for forecasting flood risk and developing mitigation strategies. This paper aims to address this gap by developing a method to quantitatively describe and analyze these dynamic spatial-temporal rainfall patterns.

Methods

The core of this study lies in the application of a manifold learning algorithm, specifically Locally Linear Embedding (LLE), to analyze high-dimensional rainfall data. The process involves:

  1. Data Representation: Rainfall data from multiple weather stations over time is structured into a high-dimensional matrix representing the spatial-temporal characteristics of rainstorm events. Each event is characterized by a proportion matrix where Hit represents the percentage of rainfall at station i at time t relative to the total rainfall across all stations at that time.
  2. Dimensionality Reduction (LLE): LLE is employed to reduce the dimensionality of this complex data space. The algorithm assumes that local linear relationships in the high-dimensional space are preserved in the low-dimensional space. This allows for the extraction of "effective" features without losing intrinsic information.
  3. Dynamic Cluster Analysis and Feature Extraction: After dimensionality reduction, a dynamic clustering algorithm is used to group similar rainstorm patterns. The mean of each cluster is then extracted as a representative feature. This process is iterated until convergence, yielding final clusters and their representative centers in both the low-dimensional and high-dimensional spaces. These extracted features are then analyzed to understand the spatial-temporal distribution characteristics.

This approach is novel in its application of manifold learning for rainstorm pattern analysis, moving beyond traditional statistical methods to capture complex, non-linear relationships.

Datasets

The study utilizes historical 5-minute rainfall data from automatic weather stations in two distinct regions:

  • Beijing: Data from 122 automatic weather stations collected from 1999 to 2020.
  • Shenzhen: Data from 63 automatic weather stations collected from 2008 to 2021.

Prior to analysis, the raw rainfall data underwent a rigorous cleaning and screening process to eliminate unreasonable values and identify effective rainfall events based on specific criteria (e.g., rainfall exceeding a threshold, spatial consistency, duration of effective rainfall).

Key Insights and Ideas

  • Topography vs. Water Vapor: A key insight is the differing drivers of rainstorm patterns in Beijing and Shenzhen. In Beijing, the spatial-temporal distribution strongly correlates with underlying topography, suggesting a more predictable influence of terrain on rainfall. In contrast, Shenzhen's rainstorm patterns are more closely linked to the source of water vapor, leading to more complex and diverse variations.
  • Manifold Learning for Dynamic Patterns: The study demonstrates that manifold learning can effectively extract dynamic spatial-temporal characteristics of rainfall that static indicators miss. This allows for a more nuanced understanding of how rainstorms develop and move.
  • Physical Mechanism Correlation: The extracted features from the machine learning algorithm were found to be consistent with known physical mechanisms, such as climatic characteristics and topographic conditions, lending credibility to the method's findings.
  • Distinct Rainstorm Types: In Shenzhen, the analysis identified three distinct types of rainstorm patterns, each with unique spatial-temporal evolution, further highlighting the complexity of rainfall dynamics in coastal regions.

Key Results

  • Rainstorm Duration: Heavy rainstorms in Beijing typically last about 24 hours, while in Shenzhen, they can extend to 72 hours (2-3 days).
  • Beijing Rainstorm Patterns: Rainstorms in Beijing tend to start in the western and northern mountainous areas and move towards the plains. The movement is relatively stable, often following a southwest-to-northeast trend, consistent with the underlying mountain-plain topography.
  • Shenzhen Rainstorm Types: Three distinct types of rainstorm patterns were identified in Shenzhen:
  • Type I: Starts from the northwest and moves southeast, with heavier rainfall in the north and northwest.
  • Type II: Originates from the southern and southeastern coastal areas and moves north, with rainfall concentrated in these coastal regions.
  • Type III: Begins on the southeast coast and moves north/northwest, with heavier rainfall in the east and southeast, and lighter rainfall in the west.
  • Physical Explanations: The identified rainstorm types in Shenzhen were explained by physical mechanisms related to cold air masses, southwest monsoons, and eastern waves/typhoons, demonstrating the method's ability to reveal interpretable patterns.

Limitations and Future Work

  • Data Limitations: The study acknowledges that while the algorithm can identify patterns, it doesn't directly explain the physical mechanisms of rainstorms. The accuracy and objectivity of the results could be improved with increased data timing and expanded data ranges.
  • Future Work: The authors suggest improving the method by incorporating meteorological data such as radar echo maps and cloud maps. Further development could enhance the accuracy of early risk identification for rainstorms. The potential for the method to analyze spatio-temporal distribution characteristics of rainfall for regional design purposes is also highlighted.

Summary updated

Abstract

Identifying the patterns of rainstorms is essential for improving the precision and accuracy of flood forecasts and constructing flood disaster prevention systems. In this study, we used a manifold learning algorithm method of machine learning to analyze rainstorm patterns. We analyzed the spatial–temporal characteristics of heavy rain in Beijing and Shenzhen. The results showed a strong correlation between the spatial–temporal pattern of rainstorms and underlying topography in Beijing. However, in Shenzhen, the spatial–temporal distribution characteristics of rainstorms were more closely related to the source of water vapor causing the rainfall, and the variation in characteristics was more complex and diverse. This method may be used to quantitatively describe the development and dynamic spatial–temporal patterns of rainfall. In this study, we found that spatial–temporal rainfall distribution characteristics, extracted by machine learning technology could be explained by physical mechanisms consistent with the climatic characteristics and topographic conditions of the region.
BibTeX
@inproceedings{liu2022spatial,
  author = {Yuanyuan Liu and Yesen Liu and Hancheng Ren and Longgang Du and Shu Liu and Li Zhang and Caiyuan Wang and Qiang Gao},
  title = {{Spatial and Temporal Pattern of Rainstorms Based on Manifold Learning Algorithm}},
  booktitle = {Water},
  year = {2022},
  doi = {10.3390/w15010037},
}
Figure