International Journal of Applied Earth Observations and Geoinformation (Nov 2022)
Enhanced all-weather precipitable water vapor retrieval from MODIS near-infrared bands using machine learning
Abstract
Four novel PWV retrieval approaches based on machine learning methods are for the first time developed to estimate all-weather precipitable water vapor (PWV) from near-infrared (NIR) measurements of the Moderate Resolution Imaging Spectroradiometer (MODIS) instrument. The four retrieval approaches are Back Propagation Neural Network (BPNN), Gradient Boosting Decision Tree (GBDT), Generalized Regression Neural Network (GRNN), and eXtreme Gradient Boosting (XGBoost). The transmittance, latitude, longitude, elevation, cloud, season, and solar zenith angle information, in association with the MODIS NIR PWV’s performance, are utilized. The in-situ one-year PWV data collected in 2017 from 453 Global Positioning System (GPS) sites in Australia and 214 GPS sites in China are utilized as target water vapor estimates for model training. Independent of the 2017 training data, two-year data observed in 2018–2019 in Australia and China are utilized to validate the four models’ performance. The results indicate that the retrieval algorithms can greatly improve the PWV retrieval accuracy from MODIS NIR observations under all-weather conditions, reducing the impact of clouds on NIR PWV retrieval. The new all-weather PWV estimates obtain R2 in the range of 0.83 ∼ 0.86, root-mean-square-error (RMSE) in the range of 4.71 mm ∼ 5.28 mm, and mean bias (MB) in the range of 0.18 mm ∼ 0.51 mm, significantly outperforming the official MODIS NIR PWV product (R2 = 0.31, RMSE = 12.03 mm, and MB = -3.04 mm). The reduction in RMSE is 60.85 % for BPNN, 59.68 % for GBDT, 56.69 % for GRNN, and 57.27 % for XGBoost. The new all-weather PWV results show a superior retrieval accuracy compared to the official MODIS NIR confident-clear PWV product, illustrating the effectiveness of the models. This could be because the retrieval models have considered multiple dependence parameters that affect the performance of MODIS-observed NIR PWV. The retrieval algorithms exhibit little spatial or temporal dependence and they can be applied to other regions and periods. This work provides a more accurate way to retrieve all-weather PWV estimates from satellite NIR measurements considering multiple dependence parameters – location, cloud, season, and solar zenith angle information.
