Xu et al. (2025) A hybrid approach for regionalization of precipitation based on maximal discrete wavelet transform and growing neural gas network clustering

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-11-18
• Authors: Tao Xu, Ben Ma, Xuan He, Ali Arshaghi
• DOI: 10.1038/s41598-025-24400-1

Research Groups

• College of Mechanical and Electrical Engineering, Wuyi University, Wuyishan, China
• The Key Laboratory for Agricultural Machinery Intelligent Control and Manufacturing of Fujian Education Institutions, Wuyi University, Wuyishan, China
• Department of Electrical Engineering, CT.C., Islamic Azad University, Tehran, Iran

Short Summary

This study developed a hybrid methodology combining Maximal Overlap Discrete Wavelet Transform (MODWT) and Growing Neural Gas (GNG) clustering to regionalize precipitation patterns in China using 45 years of monthly data from 123 stations. The approach successfully identified 12 homogeneous precipitation clusters, demonstrating improved accuracy and robustness in capturing multiscale temporal variability for water resource planning.

Objective

• To develop a novel hybrid analytical framework integrating Maximal Overlap Discrete Wavelet Transform (MODWT) and Growing Neural Gas (GNG) clustering for spatial precipitation regionalization in China.
• To improve the accuracy and robustness of spatial clustering by extracting multiscale precipitation features from MODWT output and using them as input for the GNG algorithm, thereby optimizing the spatial classification process.

Study Configuration

• Spatial Scale: China, approximately 9.6 x 10^12 square meters, utilizing data from 123 synoptic weather stations distributed across diverse climatic regions.
• Temporal Scale: Monthly precipitation time series data spanning a 45-year period (January 1980 – December 2024), resulting in 540 months of data for each station.

Methodology and Data

• Models used:
  • Maximal Overlap Discrete Wavelet Transform (MODWT) with Daubechies 4 (db4) mother wavelet for time series decomposition.
  • Shannon entropy for feature extraction from MODWT sub-series.
  • Growing Neural Gas (GNG) algorithm for unsupervised clustering.
  • Comparative clustering methods: K-means and Self-Organizing Map (SOM).
  • Evaluation criteria: Silhouette Coefficient (SC), Davies–Bouldin Index (DBI), Calinski–Harabasz (CH) Index.
  • Wavelet coherence analysis to examine relationships between precipitation and large-scale climate oscillation indices.
• Data sources:
  • Monthly cumulative precipitation data from 123 synoptic weather stations in China, obtained from the China Meteorological Administration (CMA) via the Meteostat site.
  • Large-scale climate oscillation indices: Pacific Decadal Oscillation (PDO), Pacific–North American pattern (PNA), El Niño–Southern Oscillation (ENSO), and North Atlantic Oscillation (NAO).

Main Results

• The hybrid MODWT-GNG model successfully identified 12 homogeneous precipitation clusters across China.
• The proposed model achieved a maximum Silhouette Coefficient (SC) of 0.68, indicating strong inter-cluster separation and intra-cluster compactness, which was significantly higher than clustering without MODWT preprocessing (SC = 0.56). It also showed improved Davies-Bouldin (0.91) and Calinski-Harabasz (18.98) indices.
• MODWT decomposed precipitation time series into five frequency-based sub-series (W1–W5 and V5), capturing variability across 2- to 32-month cycles.
• Shannon entropy analysis revealed the highest precipitation variability (MWE values) in the D3 (8-month) and A5 (trend) sub-series, particularly in northern and northwestern China, while eastern and southern regions exhibited more stable patterns.
• The identified clusters showed distinct hydroclimatic characteristics, aligning well with Köppen–Geiger climate classification zones (e.g., arid regions in the northwest, humid subtropical in the southeast).
• Wavelet coherence analysis indicated that the Pacific Decadal Oscillation (PDO) and North Atlantic Oscillation (NAO) indices exhibited the most persistent and coherent relationships with regional rainfall, primarily in the mid-frequency band (0.0625–0.125 cycles/sample). The Pacific–North American pattern (PNA) and El Niño–Southern Oscillation (ENSO) showed less coherent and more episodic relationships.

Contributions

• Introduces a novel and robust hybrid framework for precipitation regionalization by integrating Maximal Overlap Discrete Wavelet Transform (MODWT) for multiscale temporal feature extraction with the Growing Neural Gas (GNG) clustering algorithm.
• Demonstrates enhanced accuracy and interpretability of spatial precipitation clustering compared to traditional methods by effectively capturing complex, nonlinear, and multi-scale temporal dynamics.
• Provides a systematic methodology for delineating homogeneous precipitation regions, offering valuable insights for region-specific water resource management, climate adaptation policies, and hydrological risk assessment in China.
• Validates the physical soundness of the clustering framework by linking identified clusters to established Köppen–Geiger climate classification zones.

Funding

• Fujian Natural Science Foundation project: Application of object detection technology based on deep learning to tea impurity recognition (2024J01910).

Citation

@article{Xu2025hybrid,
  author = {Xu, Tao and Ma, Ben and He, Xuan and Arshaghi, Ali},
  title = {A hybrid approach for regionalization of precipitation based on maximal discrete wavelet transform and growing neural gas network clustering},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-24400-1},
  url = {https://doi.org/10.1038/s41598-025-24400-1}
}