the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 18 May 2026
Cross-temporal downscaling and fusion for hourly 0.01° precipitation estimation: A case study in Youxian District, China
Abstract. Reliable precipitation data are essential for fine-scale hydrological applications at the regional level. Consequently, numerous studies have sought to generate high-resolution and high-accuracy precipitation products through spatial downscaling of satellite-based precipitation estimates and bias correction using ground observations. However, few such studies have considered the sub-daily scale, which holds greater application value. In this study, a cross-temporal "downscaling-fusion" framework, termed CTDF, is proposed. Both stages employ extreme gradient boosting (XGBoost) modeling: the first stage spatially downscales daily GPM precipitation from 0.1° to 0.01° using various high-resolution environmental factors, while the second stage fuses the downscaled GPM, cloud properties, and rain gauge observations to generate the final hourly precipitation estimates. With Youxian District, China as the study area, the performance of CTDF was compared against five alternative schemes, and the spatial distribution of the generated precipitation was analyzed. Results indicate that: (1) CTDF exhibits the best overall performance (CC = 0.81, MAE = 0.88 mm, RMSE = 1.95 mm, Bias = 0.4 %), mitigating the systematic underestimation inherent in the original GPM product, while omitting either stage results in performance degradation; (2) CTDF demonstrates more robust performance across different precipitation intensities and diurnal conditions; (3) CTDF substantially enhances the representation of spatial precipitation heterogeneity, increasing the coefficient of variation (CV) of GPM by 170 % and 255 % for convective and stratiform precipitation events, respectively. Overall, the two-stage collaborative design of CTDF achieves spatial refinement and accuracy improvement, providing a viable technical pathway for generating high spatiotemporal resolution precipitation products.
- Preprint
(2777 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-1468', Anonymous Referee #1, 18 Jun 2026
-
RC2: 'Comment on egusphere-2026-1468', Anonymous Referee #2, 26 Sep 2026
This manuscript proposes a cross-temporal downscaling and fusion framework (CTDF) that integrates GPM IMERG, Himawari-8 cloud properties, environmental variables, and rain-gauge observations to generate hourly precipitation estimates at 0.01° resolution. However, several critical methodological issues remain unresolved. I therefore recommend major revision.
Major Comments
- GPM IMERG provides precipitation data at a 30-min resolution, while the Himawari-8 cloud properties are available at a 10-min resolution. Although LST is daily, NDVI is a 16-day composite, and DEM and geographic coordinates are static, these low-frequency or static variables can still be used as predictors in an hourly model. Therefore, their temporal resolutions do not necessitate aggregating IMERG to the daily scale before downscaling. The authors should explain why an hourly fusion model was not developed directly using hourly IMERG, hourly cloud properties, DEM, NDVI, LST, and geographic coordinates. The methodological necessity and advantages of introducing an intermediate daily downscaling step are not sufficiently demonstrated.
Moreover, the “GPM + cloud fusion” scheme already achieves CC = 0.79 and RMSE = 2.00 mm, compared with CC = 0.81 and RMSE = 1.95 mm for the complete CTDF framework. Thus, the improvement attributable to the first stage appears modest. The authors should distinguish its contribution to fine-scale spatial patterns from its contribution to station-level statistical accuracy and avoid overstating the synergistic effect of the two stages. - In Eqs. (3) and (4), R(h) has no spatial index. Based on the current description, it appears to be calculated from the area-total hourly and daily GPM precipitation. If the same R(h) is applied to all 0.01° grid cells on a given day, the disaggregated grids will share essentially the same diurnal evolution and differ only in their daily totals. This treatment may not adequately represent the movement of convective systems, the evolution of local precipitation centers, or temporal differences among the original GPM grid cells.
The authors should clarify:
(1) whether R(h) is calculated for the entire study area or separately for each 0.1° GPM grid cell;
(2) if an area-wide ratio is used, what its physical justification is;
(3) if grid-specific ratios are used, how spatial indices should be incorporated into the equations;
(4) whether the sum of the 24 hourly estimates for each 0.01° grid cell strictly equals its daily precipitation total; and
(5) how R(h) is calculated when the original daily GPM precipitation is zero or close to zero. - The study area covers approximately 1,020 km² and likely contains only about ten native 0.1° GPM grid cells; the exact number should be reported. Although multiple grid–day samples can be generated, the number of independent spatial locations remains very limited. It is therefore unclear whether XGBoost can robustly learn the relationships between precipitation and DEM, NDVI, LST, cloud properties, and geographic location.If dense gauge observations are unavailable outside Youxian District, the GPM-targeted first-stage model could still be trained over a larger area, while the second-stage gauge-based fusion could remain local. This would help determine whether the model captures robust precipitation–environment relationships or merely fits the characteristics of a few local GPM grid cells.
- In the first stage, samples were randomly divided into training and test sets at a ratio of 3:1. For gridded spatiotemporal data, random splitting may place samples from the same location on different dates, or neighboring samples from the same precipitation event, in both sets. This reduces test-set independence and may lead to overly optimistic estimates of model generalization. In the second stage, the gauges appear to have been randomly divided only once at a ratio of 1:1. Repeated station splits or spatial cross-validation should be conducted, with the mean performance and uncertainty ranges reported, to ensure that the results are not dependent on a particular station split.
- The independent-year evaluation uses five gauges and four precipitation events from 2023. This is a useful supplementary analysis, but the sample size is insufficient to demonstrate robust generalization across years and climatic conditions. The representativeness of the selected gauges and events is also unclear. If a more comprehensive evaluation is not possible because of data limitations, this analysis should be described as an “independent-year case study” or “preliminary cross-year assessment,” and the corresponding generalization claims should be moderated.
Minor issues:
- The complete predictor set, target variable, sample organization, and final hyperparameters should be reported separately for each XGBoost model. The relevant information is currently scattered across the preprocessing and results sections, which limits reproducibility.
- Different correlation thresholds are used for cloud-feature selection at the daily (0.30) and hourly (0.20) scales. Their justification should be provided. The authors should also clarify whether feature selection was conducted using only the training data to avoid information leakage.
- Please clarify how the categorical cloud-type variable was encoded for XGBoost and specify which variables were processed using bilinear interpolation, nearest-neighbor interpolation, and pixel averaging.
- The daytime and nighttime periods are defined in UTC. The corresponding local times and the basis for this classification should also be provided. The accumulation period and time zone used to define daily precipitation should be clearly stated.
- Several table-numbering and caption errors require correction. For example, the text refers to Table 1, whereas the following table is labeled Table 2. In addition, the caption of Table 5 (“Daily/hourly derived cloud-property features”) does not match its content. The authors should carefully check all table numbers, captions, equations, symbols, and formatting.
Citation: https://doi.org/10.5194/egusphere-2026-1468-RC2 - GPM IMERG provides precipitation data at a 30-min resolution, while the Himawari-8 cloud properties are available at a 10-min resolution. Although LST is daily, NDVI is a 16-day composite, and DEM and geographic coordinates are static, these low-frequency or static variables can still be used as predictors in an hourly model. Therefore, their temporal resolutions do not necessitate aggregating IMERG to the daily scale before downscaling. The authors should explain why an hourly fusion model was not developed directly using hourly IMERG, hourly cloud properties, DEM, NDVI, LST, and geographic coordinates. The methodological necessity and advantages of introducing an intermediate daily downscaling step are not sufficiently demonstrated.
Data sets
Cross-temporal downscaling and fusion for hourly 0.01° precipitation estimation: A case study in Youxian District, China Wei He et al. https://doi.org/10.5281/zenodo.20118493
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 264 | 94 | 41 | 399 | 31 | 32 |
- HTML: 264
- PDF: 94
- XML: 41
- Total: 399
- BibTeX: 31
- EndNote: 32
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Overall, the manuscript is well organized and the methodology is reasonably designed. The two-stage "downscaling-first, fusion-second" framework offers a degree of novelty in addressing the need for hourly, high-resolution precipitation products. The comparison against five alternative schemes, the spatial visualization results, and the cross-year (temporal generalization) validation together provide reasonably convincing evidence for the effectiveness and robustness of the proposed method. The study has practical value for applications such as regional disaster early warning, and is worthy of recognition. I offer a few questions and suggestions below for the authors' consideration during revision:
1. The Abstract does not mention the results of the independent validation of the proposed method for 2023. It is suggested that the authors add a brief summary of these results.
2. In Section 2.2.4 (Ground observation data), the quality control workflow for the rain gauges is summarized only as "outlier detection, spatial consistency check, and temporal consistency check." It is suggested that the authors provide some specific methodological description of these three steps, with appropriate references where available.
3. The temporal disaggregation in Eqs. (3)–(4), and its accompanying description, require further clarification. The source method cited by the authors, Ma et al. (2022), is computed on a per-0.1° grid basis—each grid cell uses the ratio of its own hourly value to its own daily value to obtain the allocation weight, so that the intra-day variation of each cell is preserved separately. In Eq. (4), however, both the numerator and denominator of R(h) are described as "summing all 0.1° grids in the study area," which yields a single value of R(h) per time step that is then applied to all 0.01° pixels. This is not fully consistent with the cited method. It is unclear how the authors actually implemented this step: if the calculation is in fact performed per grid cell, then Eq. (4) is not written accurately and should be revised; if a single study-area-wide ratio was indeed used, please justify this treatment and provide an appropriate supporting reference.
4. The GMWWDA method proposed by Ma et al. (2020) is used as one of the benchmark schemes in this study. In that original paper, the authors derived and compared GPM downscaling results based on several different cloud properties; however, the present manuscript does not specify which cloud property was used for the downscaling result adopted here. It is suggested that the authors briefly clarify this.
5. In Section 3.3.2 (Accuracy assessment across different precipitation intensity classes), the precipitation is divided into three intensity categories (Light, Moderate-to-heavy, and Rainstorm and above) for the inter-scheme comparison, but it is not explained how the thresholds for this classification were determined, which somewhat undermines the rigor. The authors are advised to clarify this, and to indicate the source (e.g., a relevant reference or standard) if such a basis exists.
6. There are several minor errors in the text, for example the incorrect numbering of Table 1 and the incorrect caption of Table 5. The authors are advised to check the manuscript carefully and correct these issues.