PaperlayerView exact revision on arXiv

Source-linked AI summary

A Data-Driven Approach for Accurate Rainfall Prediction

Shilpa Manandhar, Soumyabrata Dev, Yee Hui Lee, Yu Song Meng, Stefan Winkler

arXiv:1907.04816v1physics.ao-ph

TL;DR

Rain prediction from GPS-derived PWV remains challenging because rainfall depends on atmospheric factors beyond water vapor, and prior PWV-based forecasts had high false-alarm rates. The paper studies weather and seasonal/diurnal features, then uses selected predictors in a machine-learning model; it reports significantly reduced false-alarm rates compared with prior literature.

  • Problem

    PWV is a useful rain indicator, but it does not necessarily signal rainfall, and earlier PWV-based forecasts had high false-alarm rates.

    Source: §I · p. 1

  • Method

    The paper studies weather features and seasonal/diurnal variables, identifies useful predictors, and applies them in a machine-learning rainfall model.

    Source: §I · p. 1

  • Results

    The approach achieves a significant reduction in false-alarm rates compared with the literature.

    Source: §VI · p. 8

  • Takeaways & Limitations

    PWV contributes most to detection, while solar radiation, day-of-year, and hour-of-day contribute to reducing false alarms.

    Source: §VI · p. 8

Abstract

from arXiv · show

In recent years, there has been growing interest in using Precipitable Water Vapor (PWV) derived from Global Positioning System (GPS) signal delays to predict rainfall. However, the occurrence of rainfall is dependent on a myriad of atmospheric parameters. This paper proposes a systematic approach to analyze various parameters that affect precipitation in the atmosphere. Different ground-based weather features like Temperature, Relative Humidity, Dew Point, Solar Radiation, PWV along with Seasonal and Diurnal variables are identified, and a detailed feature correlation study is presented. While all features play a significant role in rainfall classification, only a few of them, such as PWV, Solar Radiation, Seasonal and Diurnal features, stand out for rainfall prediction. Based on these findings, an optimum set of features are used in a data-driven machine learning algorithm for rainfall prediction. The experimental evaluation using a four-year (2012-2015) database shows a true detection rate of 80.4%, a false alarm rate of 20.3%, and an overall accuracy of 79.6%. Compared to the existing literature, our method significantly reduces the false alarm rates.

I. INTRODUCTION

GPS-derived PWV offers good spatio-temporal resolution for rainfall monitoring, but PWV alone does not capture all rain triggers and prior forecasts had high false-alarm rates. This paper studies weather and seasonal/diurnal features to identify useful predictors for machine-learning rainfall prediction.

I. INTRODUCTION

  • GPS estimates PWV from signal delay and provides good spatio-temporal resolution in all weather conditions.The introduction contrasts GPS with radiosondes, microwave radiometers, and satellite-based retrievals.Source: §I · p. 1
  • Earlier PWV-based forecasts detected more than 80% of events but had false-alarm rates of 60–70%, motivating consideration of other atmospheric factors.High PWV does not necessarily mean rainfall will occur.Source: §I · p. 1
  • The study emphasizes identifying important rainfall predictors, their inter-dependencies, and their contributions rather than only comparing machine-learning algorithms.The introduction notes prior work combining weather features and machine learning.Source: §I · p. 1
  • The paper assesses weather and seasonal factors individually, then uses identified predictors in a machine-learning algorithm for rainfall prediction.Seasonal and diurnal variables are often neglected in related studies.Source: §I · p. 1

II. METEOROLOGICAL SENSOR DATABASE AND DATA PROCESSING

The paper describes surface weather measurements and GPS-derived column water vapor, with PWV calculated from GPS signal delays for two Singapore stations.

II. METEOROLOGICAL SENSOR DATABASE AND DATA PROCESSING

  • The database section covers surface weather parameters and total-column water vapor content.Source: §II · p. 2

A. GPS-Derived Water Vapor Content

  • GPS-derived PWV uses the zenith wet delay, which reflects atmospheric water vapor and humidity, while the hydrostatic delay contributes about 90% of total zenith delay.The wet delay is used to calculate PWV.Source: A. GPS-Derived Water Vapor Content · p. 2
  • The study processes wet-delay values with GIPSY-OASIS and calculates PWV using a dimensionless factor and liquid-water density.The factor was derived using radiosonde data from 174 stations; the selected processing uses the software's static hydrostatic-delay model.Source: A. GPS-Derived Water Vapor Content · p. 2
  • PWV data at five-minute resolution come from NTUS for 2012–2015 and SNUS for 2016.Both stations are in Singapore.Source: A. GPS-Derived Water Vapor Content · p. 2

B. Weather Station Data

Weather-station observations provide surface variables and rainfall rates alongside GPS-derived PWV; co-located sky-camera images support visual checks of atmospheric conditions.

B. Weather Station Data

  • The stations record temperature, relative humidity, dew point, solar radiation, and rainfall, with NTUS data from 2012–2015 and SNUS data from 2016.NTUS weather data are recorded every minute and SNUS data every five minutes; SNUS is used for benchmarking and validation.Source: B. Weather Station Data · p. 2
  • Co-located WAHRSIS sky-camera images provide visual validation, and NTUS weather variables are sampled every five minutes to match GPS-PWV timing.Source: B. Weather Station Data · p. 2

III. APPROACH & TOOLS

The approach uses supervised SVM classification to distinguish rain from no-rain cases using weather features, then evaluates predictions against observations with detection, false-alarm, accuracy, and missed-detection metrics.

A. Supervised Machine Learning Technique

  • Rainfall prediction is framed as supervised classification of samples labeled rain or no-rain, using SVM with weather parameters as features.The method is described as computationally efficient, and training samples define a maximum-margin separator.Source: A. Supervised Machine Learning Technique · pp. 2–3

B. Evaluation Metrics

  • The confusion matrix compares predicted output with observed outcomes and distinguishes true positives, true negatives, false positives, and false negatives.Source: B. Evaluation Metrics · p. 3, Table I · p. 3
  • Evaluation reports true detection, false alarm, accuracy, and missed detection rates, defined from the confusion-matrix counts.Source: B. Evaluation Metrics · p. 3

C. Downsampling Technique

The paper combines seasonal and diurnal features with weather measurements, while balancing rain and no-rain samples to address the database’s strong class imbalance.

C. Downsampling Technique

  • In 2015 NTUS data, only 5,017 of 525,120 valid points contained rain, a minority-to-majority ratio of nearly 1:104.The paper notes that training on this skewed dataset could bias the model toward non-rain cases and compromise generalization.Source: C. Downsampling Technique · p. 3
  • Downsampling retains all rain cases and randomly selects no-rain cases to balance the labels; without it, accuracy dropped and confidence intervals widened significantly.The paper notes that 1:1 is common, though other ratios are possible.Source: C. Downsampling Technique · p. 3

A. Features

  • The selected weather features are temperature, relative humidity, solar radiation, dew point temperature, and PWV, alongside seasonal and diurnal features.The paper uses day-of-year to represent seasonality and hour-of-day to represent diurnal effects.Source: A. Features · p. 3
  • Singapore’s monsoon seasons have differing rainfall patterns, motivating day-of-year as a feature for seasonal effects.The paper names four seasons: North-East, First-Inter, South-West, and Second-Inter monsoons.Source: A. Features · p. 3
  • Because hour-of-day and day-of-year repeat cyclically, each is represented by sine and cosine components for use in methods that include these features.The cycle repeats every 24 hours for hour-of-day and every 365 days, or 366 in leap years, for day-of-year.Source: A. Features · pp. 3–4

B. Feature Correlation

The correlation analysis finds strong relationships among several surface weather variables, while PWV is largely distinct and cyclic time features expose seasonal and diurnal relationships.

B. Feature Correlation

  • Temperature correlates strongly with relative humidity (around −0.9), and positively with solar radiation (0.6) and dew point temperature (0.5).The paper relates the patterns to daytime increases in solar radiation and temperature alongside falling relative humidity.Source: B. Feature Correlation · p. 4
  • PWV has no strong correlation with the other features, apart from a small positive correlation with relative humidity; temperate regions show stronger RH–PWV correlation.The paper attributes the temperate-region difference to wider temperature ranges across seasons and locations.Source: B. Feature Correlation · p. 4
  • Sine and cosine components of day-of-year and hour-of-day correlate with temperature, relative humidity, dew point, and solar radiation.The paper says this cyclic representation reveals relationships that direct hour- and day-of-year values had underestimated.Source: B. Feature Correlation · p. 4
  • Only relative humidity and temperature have very high correlation in the analysis, suggesting these weather variables can each contribute to a weather phenomenon.The paper then examines all the weather variables with respect to rain events.Source: B. Feature Correlation · p. 4

C. Time Series Observation

The time-series observations show that PWV rises before rainfall in both daytime and nighttime examples, while surface-variable and solar-radiation changes depend on timing and conditions.

C. Time Series Observation

  • PWV increases within three hours before the two daytime rain events in window W1, consistent with a reported tropical-region pattern.The increases begin around 11:00 before 14:00 rain and around 15:00 before 18:00 rain.Source: C. Time Series Observation · p. 4
  • During W1 rain, temperature drops toward the dew point, relative humidity reaches 90–100%, and solar radiation decreases as rain clouds block the sun.The WAHRSIS images at 14:05 and 17:30 show thick dark clouds.Source: C. Time Series Observation · pp. 4–5
  • Window W2 shows the same pattern: PWV rises before rain, temperature falls to the dew point, relative humidity reaches 100%, and solar radiation decreases.PWV begins rising around 12:00 before the 15:30 rain event; clouds are visible in the 15:15 image.Source: C. Time Series Observation · p. 5
  • At night, temperature, relative humidity, dew point, and solar radiation show little variation, but PWV rises about three hours before the midnight rain event in W3.The paper also summarizes that PWV and solar radiation change before daytime rain, whereas nighttime pre-rain changes are distinct for PWV.Source: C. Time Series Observation · p. 5

V. RAINFALL PREDICTION

The prediction setup uses 30-minute feature segments to predict rain after a five-minute lead time, combining meteorological, seasonal, and diurnal information in one model.

V. RAINFALL PREDICTION

  • Each prediction example uses 30 minutes of data to determine whether rain occurs after a five-minute lead time, with rainfall within a 30-minute window counted as one event.The paper uses the same prediction scenario as its earlier work but studies combined meteorological features with PWV.Source: §V · p. 6
  • Unlike the earlier work’s separate seasonal models, this paper combines seasons into a single model using seasonal and diurnal features.The stated evaluation uses 2012–2015 NTUS data and 2016 SNUS data.Source: §V · p. 6

A. Assessment of Individual Features

Feature usefulness differs between classifying rain and predicting it: PWV and solar radiation stand out for prediction, motivating evaluation of seasonal and diurnal features alongside weather variables.

A. Assessment of Individual Features

  • All tested features distinguish rain from no-rain conditions during the day, but not all perform well for prediction.Most features also classify effectively at night; solar radiation is reported only for daytime because it is absent at night.Source: A. Assessment of Individual Features · p. 6
  • PWV separates true detections from false alarms for prediction in both day and night, while solar radiation gives the highest daytime true detection and is not useful at night.Temperature, relative humidity, and dew point can distinguish rain from no-rain conditions but change during rain rather than beforehand.Source: A. Assessment of Individual Features · p. 6

B. Selection of Optimal Features

1) Adding Features:

The feature-addition and removal experiments identify PWV, solar radiation, day of year, and hour of day as the strongest combination for prediction, with distinct roles in detection and false-alarm control.

1) Adding Features:

  • Adding SR to PWV, DoY, and HoD raises TD from 79.8% to 80.4%, lowers FA from 31.9% to 20.3%, and increases accuracy from 68.2% to 79.6%.These results use a 5-minute prediction lead time.Source: 1) Adding Features · p. 7
  • With all features, TD is 83.0%, FA is 20.6%, and accuracy is 79.4%; adding T, RH, and DPT raises TD but also raises FA.The highest accuracy is reported for PWV, SR, DoY, and HoD rather than the all-feature pool.Source: 1) Adding Features · p. 7

2) Removing Features:

  • Removing PWV significantly lowers TD, while removing DoY and HoD raises FA by almost 8.5% and removing SR raises FA from 20.6% to 29.4%.Removing DPT, T, or RH slightly lowers TD while leaving FA and accuracy unchanged from the all-feature result; these features do not contribute much to prediction.Source: 2) Removing Features · p. 7
  • PWV helps maintain high TD, whereas SR, DoY, and HoD help control false alarms; removing DPT, RH, and T can slightly reduce false alarms.The authors relate the seasonal and diurnal feature findings to their earlier work.Source: 2) Removing Features · p. 7
  • 80.4% TD, 20.3% FA, and 79.6% accuracy are achieved with PWV, SR, DoY, and HoD, the best-performing feature combination.The reported TD corresponds to a missed detection rate of around 19.6%.Source: 2) Removing Features · p. 8

C. Benchmarking

The paper benchmarks its PWV, SR, HoD, and DoY approach against prior methods and reports lower false-alarm rates. Its conclusions assign PWV a role in detection and SR, DoY, and HoD roles in reducing false alarms.

C. Benchmarking

  • The proposed approach significantly improves TD and FA over,, and; against, average FA falls by 16.2 percentage points with only a small TD reduction.The comparison covers 2012–2015 and also reports reduced FA for SNUS GPS and co-located weather-station data.Source: Table II · p. 8, C. Benchmarking · p. 8, Table III · p. 8

VI. CONCLUSIONS & FUTURE WORK

  • PWV contributes most to achieving a high detection rate, while SR, DoY, and HoD contribute to reducing false-alarm rates.All features play a role in rainfall classification, but this subset also shows potential for rainfall prediction.Source: §VI · p. 8
  • Future work will examine feature choices across different stations with a larger dataset, as well as derivatives, additional weather features, and longer prediction lead times.Source: §VI · p. 8
Loading 1907.04816v1…