Gradient boosting machine learning to improve satellite-derived column water vapor measurement error

Area of Science:

• Atmospheric Science
• Remote Sensing
• Data Science

Background:

• Multi-Angle Implementation of Atmospheric Correction (MAIAC) algorithm provides daily column water vapor (CWV) at 1 km resolution from MODIS instruments.
• Previous work demonstrated machine learning (XGBoost) can enhance aerosol optical depth (AOD) retrieval.
• MAIAC CWV is generally well-validated, but machine learning's potential for CWV improvement remains unexplored.

Purpose of the Study:

• To assess if machine learning approaches can further improve column water vapor (CWV) estimates from the MAIAC algorithm.
• To quantify and correct measurement errors in MAIAC CWV data using machine learning.
• To evaluate the effectiveness of machine learning corrections on independent validation datasets.

Main Methods:

• Developed an extreme gradient boosting (XGBoost) model using nine features from land use, date, and ancillary MAIAC variables.
• Employed a novel spatiotemporal cross-validation approach to prevent overfitting.
• Validated the corrected MAIAC CWV against AERONET sun photometer and SuomiNet GPS network data.

Main Results:

• The XGBoost model significantly reduced root mean square error (RMSE) in MAIAC CWV: 26.9% (Terra) and 16.5% (Aqua) versus AERONET.
• Machine learning interpretation tools revealed complex error patterns and a worsening positive bias in MAIAC Terra CWV during recent summers.
• Corrections decreased RMSE by 19.7% (Terra) and 9.5% (Aqua) on independent SuomiNet GPS data.

Conclusions:

• Machine learning, specifically XGBoost, offers a viable postprocessing method to correct measurement errors in satellite-derived CWV.
• Empirically correcting MAIAC CWV data enhances its reliability and usability for Earth science and remote sensing applications.
• This approach presents an opportunity to improve the quality of global CWV datasets derived from satellite observations.