Scholarlink

<p>High-Impact Weather (HIW) events, particularly high-impact heavy rainfall, pose significant risks to urban infrastructure in Australia. Traditional forecasting approaches often struggle to resolve the complex, non-linear thermodynamic interactions that drive these infrequent events, while standard supervised machine learning models are hindered by severe class imbalance. This study presents a novel, multi-disciplinary framework that integrates synoptic climatology with unsupervised anomaly detection to classify and predict high-impact heavy rainfall events in Darwin, Sydney, Brisbane, and Perth. Using daily meteorological observations (2024–2025), we developed a multi-phase analytical framework comprising precursor, thermodynamic, kinematic, and system evolution phases to isolate the physical signatures of storm genesis. Exploratory analysis using “Danger Rose” polar histograms revealed a strong anisotropic risk pattern, with heavy rainfall predominantly associated with South-South-East (SSE) and West-South-West (WSW) vectors. Bivariate Kernel Density Estimation (KDE) revealed a distinct “Thermodynamic Lock-in” mechanism, where severe events are confined to narrow regimes of low pressure (&lt; 1010 hPa), high humidity (&gt; 60%), and compressed diurnal temperature ranges. To address the limited representation of severe events data (12.1%), we benchmarked five unsupervised anomaly detection algorithms. The results indicate that DBSCAN (Density-Based Spatial Clustering) yields the optimal performance (F1-Score: 0.319; Recall: 67.5%), significantly outperforming Isolation Forest and PCA. Topological validation via t-SNE projection confirms that high-impact heavy rainfall events form dense, cohesive clusters within the phase space rather than appearing as randomly distributed stochastic outliers. These findings prove that hybridizing physical phase-space analysis with density-based machine learning offers a robust pathway for early warning systems in data-sparse environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A Physics-guided Unsupervised Learning Framework for High-impact Heavy Rainfall Prediction in Data-sparse Environments

  • 期刊论文

  • 发表日期：

    2026年5月22日

查看全文
  • K. Karthick,
  • S K Aruna,
  • S. Krishnan,
  • S. Ravivarman,
  • R. Manikandan

摘要

High-Impact Weather (HIW) events, particularly high-impact heavy rainfall, pose significant risks to urban infrastructure in Australia. Traditional forecasting approaches often struggle to resolve the complex, non-linear thermodynamic interactions that drive these infrequent events, while standard supervised machine learning models are hindered by severe class imbalance. This study presents a novel, multi-disciplinary framework that integrates synoptic climatology with unsupervised anomaly detection to classify and predict high-impact heavy rainfall events in Darwin, Sydney, Brisbane, and Perth. Using daily meteorological observations (2024–2025), we developed a multi-phase analytical framework comprising precursor, thermodynamic, kinematic, and system evolution phases to isolate the physical signatures of storm genesis. Exploratory analysis using “Danger Rose” polar histograms revealed a strong anisotropic risk pattern, with heavy rainfall predominantly associated with South-South-East (SSE) and West-South-West (WSW) vectors. Bivariate Kernel Density Estimation (KDE) revealed a distinct “Thermodynamic Lock-in” mechanism, where severe events are confined to narrow regimes of low pressure (< 1010 hPa), high humidity (> 60%), and compressed diurnal temperature ranges. To address the limited representation of severe events data (12.1%), we benchmarked five unsupervised anomaly detection algorithms. The results indicate that DBSCAN (Density-Based Spatial Clustering) yields the optimal performance (F1-Score: 0.319; Recall: 67.5%), significantly outperforming Isolation Forest and PCA. Topological validation via t-SNE projection confirms that high-impact heavy rainfall events form dense, cohesive clusters within the phase space rather than appearing as randomly distributed stochastic outliers. These findings prove that hybridizing physical phase-space analysis with density-based machine learning offers a robust pathway for early warning systems in data-sparse environments.