[Paper Review] A Data-driven Approach to Detecting Precipitation from Meteorological Sensor Data

Shilpa Manandhar, Soumyabrata Dev+3 more|arXiv (Cornell University)|May 4, 2018

Precipitation Measurement and Analysis8 references3 citations

TL;DR

This paper proposes a machine learning approach using surface meteorological sensors and GPS-derived precipitable water vapor (PWV) to detect rainfall, achieving a high true detection rate of 87.4% and a moderate false alarm rate of 32.2% by leveraging SVM with random downsampling to address class imbalance in tropical rainfall data.

ABSTRACT

Precipitation is dependent on a myriad of atmospheric conditions. In this paper, we study how certain atmospheric parameters impact the occurrence of rainfall. We propose a data-driven, machine-learning based methodology to detect precipitation using various meteorological sensor data. Our approach achieves a true detection rate of 87.4% and a moderately low false alarm rate of 32.2%.

Motivation & Objective

• To improve rainfall detection accuracy in tropical climates using multi-source meteorological sensor data.
• To address the challenge of highly imbalanced rainfall datasets where rain events are rare.
• To evaluate the effectiveness of combining surface weather parameters and GPS-derived precipitable water vapor (PWV) for rainfall classification.
• To reduce false alarms in precipitation detection using machine learning and data balancing techniques.

Proposed method

• The study uses surface weather parameters—temperature, relative humidity, dew point, solar irradiance, and rainfall rate—collected at 1-minute resolution.
• GPS-derived precipitable water vapor (PWV) is calculated from zenith wet delay using equations involving latitude, day-of-year, and station altitude.
• Seasonal (day-of-year) and diurnal (time-of-day) features are included to capture monsoon and daily rainfall patterns in Singapore.
• A Support Vector Machine (SVM) classifier is trained on a balanced dataset using random downsampling to equalize rain and no-rain samples.
• The model is trained on 20% of data from 2010–2011 and tested on unseen data from 2012, with performance evaluated via true detection and false alarm rates.
• The dataset is resampled to a 1:1 minority-to-majority ratio to mitigate bias from the extreme class imbalance (1:70 ratio of rainy to non-rainy points).

Experimental results

Research questions

• RQ1 Can a combination of surface weather parameters and GPS-derived PWV improve rainfall detection accuracy in tropical regions?
• RQ2 How does random downsampling affect model performance in highly imbalanced rainfall detection tasks?
• RQ3 What is the impact of including diurnal and seasonal features on rainfall classification accuracy?
• RQ4 How do true detection and false alarm rates compare to prior work using only PWV data?

Key findings

• The proposed method achieves a true detection rate of 87.4% on test data and 85.8% on 2012 data, significantly outperforming prior methods in detection capability.
• The false alarm rate is reduced to 32.2% on test data and 28.5% on 2012 data, representing a notable improvement over previous studies reporting 60% false alarms.
• Random downsampling significantly improves model performance, with a 5% increase in true detection and over 12% reduction in false alarm rate compared to non-downsampled training.
• The inclusion of diurnal and seasonal features enhances model sensitivity to time-dependent rainfall patterns, particularly in monsoon seasons.
• The model maintains strong generalization on out-of-sample data from 2012, confirming robustness across years.
• The results demonstrate that combining surface sensors with GPS-derived PWV via SVM yields a more accurate and reliable rainfall detection system than PWV-only approaches.