Modify the Accuracy of MODIS PWV in China: A Performance Comparison Using Random Forest, Generalized Regression Neural Network and Back-Propagation Neural Network

Abstract

Water vapor plays an important role in climate change and water cycling, but there are few water vapor products with both high spatial resolution and high accuracy that effectively monitor the change of water vapor. The high precision Global Navigation Satellite System (GNSS) Precipitable Water Vapor (PWV) is often used to calibrate the high spatial resolution Moderate−resolution Imaging Spectroradiometer (MODIS) PWV to produce new PWV product with high accuracy and high spatial resolution. In addition, the machine learning method has a good performance in modifying the accuracy of MODIS PWV. However, the accuracy improvement of different machine learning methods and different modeling timescale is different. In this article, we use three machine learning methods, namely, the Random Forest (RF), Generalized Regression Neural Network (GRNN), and Back−propagation Neural Network (BPNN) to calibrate MODIS PWV in 2019, at annual and monthly timescales. We also use the Multiple Linear Regression (MLR) method for comparison. The root mean squares (RMSs) at the annual timescale with the three machine learning methods are 4.1 mm (BPNN), 3.3 mm (RF), and 3.9 mm (GRNN), and the average RMSs become 2.9 mm (BPNN), 2.8 mm (RF), and 2.5 mm (GRNN) at the monthly timescale. Those results are all better than the MLR method (5.0 mm at the annual timescale and 4.6 mm at the monthly timescale). When there is an obvious variation pattern in the training sample, the RF method can capture the pattern to achieve the best results since the RF achieves the best performance at the annual timescale. Dividing such samples into several sub−samples each having higher internal consistency could further improve the performance of machine learning methods, especially for the GRNN, since GRNN achieves the best performance at the monthly timescale, and the performance of those three machine learning methods at the monthly timescale is better than that of annual timescale. The spatial and temporal variation patterns of the RMS values are significantly weakened after the modeling by machine learning methods for both three methods.

PWV Inversion Model Based on Random Forest and the Trend of Its Conversion Rate with Precipitation in Hubei from 1960 to 2020

“…where p is the pressure in Pa; and q is the specific humidity in g/g. The sources of the GNSS PWV modelling data used in this paper were consistent with those mentioned in the literature [17], which were provided by China Meteorological Administration. The GNSS observation data were solved by the GAMIT software to obtain the GNSS ZTD (zenith total delay), which consisted of two parts: the zenith hydrostatic delay (ZHD) and the zenith wet delay (ZWD).…”

“…RF can effectively solve the overfitting problem and does not require feature selection. The studies in [16,17] use various machine learning algorithms including RF to model PWV-related elements and show that RF can achieve better results than the backpropagation neural network (BPNN) and the generalized regression neural network (GRNN) with a shorter training time, so RF is used for the modelling in this paper.…”

Machine Learning Model for Sentiment Analysis of COVID-19 Tweets

“…Then, these trees are voting to choose the best classification tree. The RF algorithm is considered to be un-biased, since each tree works on a sample of data, and the method works well with missing data [30], [31]. We applied RF for sentiment classification and conducted extensive experiments to optimize its performance using grid search optimization.…”

Reproducing High Spatiotemporal Resolution Precipitable Water Distributions Using Numerical Prediction Data

“…Currently, methods commonly used to observe PW amounts include radiosonde observations [10], microwave radiometers [11,12], global navigation satellite system (GNSS) [13][14][15], and satellite remote sensing [16][17][18][19]. Of these, radiosonde observations are the traditional observation method [20].…”

“…Furthermore, progress in satellite remote sensing technologies has made it possible to observe water vapor in the atmosphere on a large scale with high frequency. For example, the data collected by the MODIS sensors mounted on the Terra and Aqua satellites can provide atmospheric water vapor products with a high spatial resolution [16,19], but the disadvantage is that the temporal resolution is low. Moreover, PW amounts can be obtained at a temporal resolution of 1 h from the numerical predictions of global models, but these involve the disadvantage of having low spatial resolutions and accuracy [22,23].…”