Comparative Analysis of Random Forest, BiLSTM, and Isolation Forest for Rainfall Data Anomaly Detection to Support Automatic Rain Gauge (ARG) Performance Monitoring
TL;DR
This study aims to evaluate and compare the performance of the Random Forest, Bidirectional Long Short-Term Memory (BiLSTM), and Isolation Forest algorithms in detecting anomalies in rainfall data, as well as to implement these methods in a web-based monitoring system.
Abstract
Accuracy and reliability of rainfall data form an essential foundation for meteorological observation activities. However, the presence of missing values or unreasonable readings can degrade data quality, necessitating anomaly detection methods to support Automatic Rain Gauge (ARG) performance monitoring. This study aims to evaluate and compare the performance of the Random Forest, Bidirectional Long Short-Term Memory (BiLSTM), and Isolation Forest algorithms in detecting anomalies in rainfall data, as well as to implement these methods in a web-based monitoring system. Historical rainfall data from the ARG at the STMKG Engineering Station (STA9003) spanning 2020–2022 with a 10-minute temporal resolution were used in this research. The workflow includes data preprocessing, reference label generation using Rule-Based Quality Control, feature engineering, and model development using the three selected algorithms. Model performance was assessed using accuracy, precision, recall, F1-score, and the Area Under the ROC Curve (ROC-AUC). The results show that Random Forest achieved the best overall performance with an accuracy of 84.50% and an ROC-AUC of 0.760, while BiLSTM and Isolation Forest yielded lower performance metrics on this dataset. The detection outputs were subsequently visualized in a dashboard within a web-based monitoring system to facilitate anomaly tracking and systematic ARG performance evaluation.