| 02 Sep 2026
Accuracy assessment of MODIS water vapor products in arid regions based on ground-based GNSS observations
Abstract. The MODIS precipitable water vapor (PWV) product has been widely used in hydrometeorological and hydro-ecological research because of its kilometer-scale spatial resolution, daily temporal resolution, and global coverage. However, its regional applicability, error characteristics, and correction methods remain insufficiently evaluated in arid regions because radiosonde (RS) stations are sparsely distributed and their routine observations at fixed times rarely coincide with MODIS overpasses. Based on RS observations and two Global Navigation Satellite System (GNSS) meteorological stations over the arid Hexi Corridor, we established a two-stage method in which fixed-time RS observations were first used to identify the most suitable GNSS-PWV retrieval scheme and correct the resulting estimates, after which the corrected GNSS-PWV series was used to evaluate MODIS-PWV at satellite overpass times. Results show that the GNSS-PWV retrievals were sensitive to weighted mean temperature (Tm) estimation, and those obtained using integrated Tm agreed better with RS-PWV. After RS-based linear correction, the corrected GNSS-PWV achieved R = 0.97 and RMSE = 2.30 mm. Relative to the corrected GNSS-PWV, MODIS-PWV showed a systematic overestimation, with mean biases of 2.21 mm at the reference station GSIA and 2.88 mm at the transfer-test station GSCM, and larger biases during high-PWV periods in summer. A univariate linear correction can effectively reduce the bias, with the RMSE decreasing from 3.47 to 2.05 mm at GSIA and from 5.22 to 3.58 mm at GSCM. This two-stage RS–GNSS–MODIS method provides a methodological reference for evaluating satellite-derived water vapor products and supporting their regional application in arid regions.
This study uses one year of MODIS NIR PWV, radiosonde observations, and observations from two GNSS stations in the Hexi Corridor in China, an arid region. The objective is to assess and correct MODIS PWV data for this specific arid region, where the accuracy of MODIS NIR PWV retrievals may be affected by specific regional conditions. A two-step approach is used. First, the GNSS PWV is corrected against radiosonde data at one location near one of the GNSS stations, while also comparing 4 combinations of PWV retrievals from GNSS ZWD. Second, the corrected GNSS PWV datasets are used to assess and correct the MODIS PWV dataset.
General
The study addresses a relevant issue in satellite remote sensing PWV applications and is within the scope of AMT. Regional validation and correction studies of satellite PWV are valuable, as they can provide insights into its regional applicability and help to identify limitations and quantify accuracy not necessarily obvious from large-scale validation studies.
Given that there are only two GNSS stations and one location with RS launches in the study region, I appreciate the two-step approach taken, as well as testing the 4 combinations for GNSS PWV retrievals, which show to have significantly different impacts on the GNSS PWV time series. Transferability is tested, and performances for uncorrected and corrected datasets are compared, although this is not an independent validation of the correction method. The structure of the manuscript is clear. In the title, “arid regions” should be “arid region” since the study focuses on only one arid region in China.
Questions/comments related to the methodology and performance assessments:
Major comment: While the authors speak of corrected datasets throughout the manuscript, I think there should be a clearer distinction made between accuracy assessment and linear correction (calibration applied to the full valid dataset). The MODIS and GNSS PWV datasets and errors are compared before and after corrections, but the linear correction itself is not independently validated because it is always applied to the full valid dataset. For MODIS PWV, a zero bias is obtained after correction. This demonstrates improved agreement with the reference dataset within the same time period, but does not tell us how robust the correction method is for unseen data. An independent validation (e.g. other time periods of the dataset) of the corrected MODIS PWV is missing. This is also true for the GNSS correction in the first step. If the number of matchups is limited, a straightforward cross-validation approach, by applying linear correction to various (temporal) subsets of the data, can be used, and performance for corrected and uncorrected unseen data subsets can be compared. Also, it is not clear whether there was only one year of measurements available at the GNSS stations or whether it was selected by the authors. Otherwise, the study might benefit from extending the time series for independent validation. References to these measurement sites and datasets are missing in the text.
Also, more methodological description and result discussion on the quality control, collocation strategy, and representativeness should be included:
What is the distance between the GNSS station and RS launch site?
What is the allowed RS horizontal drift and time duration during ascent?
What are the elevations of the GNSS stations? And compared to the surroundings? Is the terrain relatively flat?
How sensitive are the performance results to different collocation approaches, like the choice of the MODIS spatial window and GNSS temporal window?
Why are there only hourly GNSS PWV measurements, while in the text it is stated that the sampling interval of the (raw) observations at the GNSS stations is 30 seconds? Do you have an idea of how variable PWV can be on such time scales for that region?
With the 3-sigma approach, outliers can be removed, but it is not clear whether these outliers are related to collocation or retrieval issues. Alternatively, one could make the matchup selection stricter in the collocation step beforehand based on predefined criteria for a better understanding of their impact on performance.
What is the impact of including probably clear pixels in the averaging of the MODIS PWV around the GNSS station? Is there a pixel buffer zone around cloudy pixels taken for the MODIS PWV field? To avoid issues with, e.g., cloud shadows? What about aerosols affecting the MODIS NIR PWV retrievals?
Several potential error/uncertainty sources are discussed for MODIS NIR PWV retrievals, related to the surface reflectance, viewing geometry, cloud contamination etc., but no effort is made to assess whether remaining errors might be consistent with known issues of MODIS NIR PWV retrievals. For example, relation to valid pixel fraction within a spatial window or viewing geometry.
2.2.1: Reference to a document describing TEQC is missing
2.2.2 Suggestion to mention local times of 00.00 and 12.00 UTC for that region. To get an idea of where in the daily cycle the measurements are performed.
2.2.3 Reference to Saastamoinen model is missing
2.2.4 Why only Terra and not Aqua? Terra overpass times?
2.3.1 It is called widely used processing methods for ZTD, but references for PPP and GAMIT are missing