the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 26 Mar 2026
Machine Learning vs. Conventional Methods for X-Band Radar Rainfall Estimation in Cyprus
Abstract. Polarimetric X-band radars offer high-resolution precipitation observations that are often challenged by attenuation, calibration errors, and absence of routine correction procedures, which limit reliable quantitative precipitation estimation (QPE). This study proposes a two-stage machine learning framework for estimating near-surface rainfall from the Cyprus national X-band radar network. In the first stage (Stage 1), feedforward neural networks correct raw ground radar reflectivity using volume-matched Ku-band measurements from the Global Precipitation Measurement (GPM) Mission dual-frequency precipitation radar (DPR). In the second stage (Stage 2), the corrected reflectivity is used as input to regression models, including support vector regression (SVR) and neural networks, to estimate rainfall rates using tipping-bucket rain gauge data. Results show that the Stage 1 networks substantially improve ground radar reflectivity, while Stage 2 SVR models outperform traditional Z–R relationships in predicting rainfall, despite residual underestimation and moderate accuracy. The study highlights the potential of machine learning methods for X-band radar QPE in environments with limited calibration and emphasizes the benefit of combining multiple radar datasets to improve spatial consistency. These findings provide practical insights for enhancing rainfall estimation in Cyprus and other regions with similar radar network constraints.
- Preprint
(5929 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
- RC1: 'Comment on egusphere-2026-445', Anonymous Referee #1, 19 Apr 2026
-
RC2: 'Comment on egusphere-2026-445', Anonymous Referee #2, 28 Apr 2026
REVIEW OF: “Machine Learning vs. Conventional Methods for X-Band Radar Rainfall Estimation in Cyprus”
The paper investigates how machine learning can improve rainfall estimates from an X-band weather radar, a system particularly sensitive to attenuation and calibration uncertainties. The authors design a two-step approach in which satellite radar observations are first used to adjust ground-based radar reflectivity measurements, aiming to reduce systematic biases and inconsistencies. In a second step, the adjusted radar signal is converted into surface rainfall using data-driven models trained on rain gauge measurements, thereby replacing traditional empirical relationships with learned nonlinear mappings. The methodology is tested over Cyprus, where the performance of different machine learning techniques is compared with standard reflectivity–rainfall conversions.
I have some major concerns regarding both the editorial and technical aspects.
- I often found a lack of clarity in the prose that tends to obscure key methodological details. In a study like this, the exact structure of the input datais essential for interpreting the results and assessing reproducibility. For each model, it should be explicitly stated what the input tensor looks like (e.g., number of features, whether vertical profiles or single levels are used, spatial/temporal context, normalization). Without this, it is impossible to understand what the networks are actually learning or to compare fairly with alternative approaches.
- Similarly, the absence of a clear description of the volume matching procedure between ground radar and GPM DPRis a major gap. This step is non-trivial and can strongly influence the results due to differences in resolution, viewing geometry, and sampling volumes.
- The two-step design is introduced more as a procedural choice than as a well-motivated strategy. The authors should clearly explain whythe problem is decomposed into (i) satellite-to-ground radar adjustment and (ii) radar-to-rainfall estimation, instead of, for example, a direct mapping from satellite observations to surface rainfall. A convincing motivation could be based on ideas such as domain adaptation, error separation (instrumental vs. microphysical), or leveraging higher-quality supervision (gauges) in an intermediate step—but none of these are articulated.
- The same applies to the change in the number of layers for the models used for the two radars; architectural decisions should be motivated (even briefly) in terms of the complexity of the mapping, input structure, or validation performance. Otherwise, it raises concerns about ad-hoc tuning or a lack of systematic model selection.
- Operations like the ones you mention (lines 199–201) appear as “black-box” preprocessing steps, and it is hard for the reader to assess the impact of these choices on the results.
- The overall quality of the figures should be improved. In particular, Figure 6 appears to contain visible artifacts that are not discussed in the text, making it difficult to determine whether they arise from the data, the processing chain, or the visualization itself. These features might be misleading and should either be explained or removed through improved plotting or preprocessing. In addition, the scatter plots are difficult to interpret due to overplotting and poor visual contrast; important structures (e.g., biases, spread, and conditional behavior) are not easily discernible.
Additional comments:
- Line 15: What do you mean by weather radar? I would be more precise about which frequency you mean.
- Line 32: cite a reference to support this statement
- Line 32: “provide high spatial resolution” better restate as “provide data at high spatial res... “or “that can perform high res observations”
- Line 75 - 77: ground radar reflectivity: is this a map obtained from scans? Or a profile? Unclear. Why not plotting it, to introduce your input data, with a scheme?
- Line 80: Reflectivity data is too generic: What type of data? Scans? Volumes? Profiles? What is the dimensionality? Unclear.
- Line 86-88: I think a figure would help to clarify which data you are considering.
- Line 89-92: What is the dimensionality of the GPM data you use?
- Line 130: How much data did you collect in the end?
- Line 143: What is the corresponding range to the ground radar? Unclear. How do you perform the volume matching? Showing some graphics to make people understand would help.
- Line 145: “corresponding”… to what?
- Line 150-151: I don't understand. The output of the first network is the volume-matched Ku reflectivity, while the input to the second is called ground reflectivity. Are these two the same? If yes, why use two different names?
- General comment: you might clarify for a non-expert, what pulse data is, what the pulse tip is.
- What is the criterion with which you decide the number of layers in the MPL?
- Line 330: You do not show any discussion on the choice of the optimal radius. If this is so relevant, then more info should be provided
- Line 337-338: You do not explain why.
- Line 159: Clarify what you mean by “lowest radar sweep.”
- Line 167-170: Why do you need a second stage? Couldn't you retrieve rain rates in the first stage? What's the added value and the motivation to add the second step?
- Line 199-201: Totally unclear what all these manipulations are needed for.
- Line 205-206: You said this already in your methodology section.
- Section 3.2.2: Why do you need to define a specific model for the PFO site, which is different from the ones defined in the previous methodology section?
- Section 3.3: I suggest putting these definitions in the appendix.
- Line 235: missing “n”
- Line 249: slight underestimation by whom? need to specify in the text... slight means what? Can you quantify it?
- Line 253: How did it result in this radius value? Can you show the procedure? It seems this is a relevant step from your discussion.
- Fig 6: There seem to be unrealistic features. Do you know what is causing them? For light green values
- Line 285: What is the criterion with which you decide the number of layers in the MPL?
- Line 330: You do not show any discussion on the choice of the optimal radius. If this is so relevant, then more info should be provided
Citation: https://doi.org/10.5194/egusphere-2026-445-RC2
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 501 | 204 | 43 | 748 | 31 | 47 |
- HTML: 501
- PDF: 204
- XML: 43
- Total: 748
- BibTeX: 31
- EndNote: 47
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
Machine Learning vs. Conventional Methods for X-Band Radar Rainfall Estimation in Cyprus
The manuscript presents a two-step machine learning framework for estimating precipitation from X-band radar observations in Cyprus. In the first step, a neural network is used to correct ground-based reflectivities using DPR data, and in the second step, the corrected reflectivities are used to estimate precipitation rates.
While the general problem of inferring precipitation rates from ground-based X-band radar measurements is relevant for any region where they are deployed, the current study exhibits fundamental shortcomings in both methodology and results. Most importantly, the reported performance metrics raise serious concerns about the validity of the proposed approach and its implementation. The coefficients of determination (r^2) reported for one of the PFO radar from 0.078 to -67.66. Negative values of r2^2 indicate that the model performs substantially worse than a trivial baseline indicating either implementation or data errors. This strongly suggests that the model, in its current form, does not capture a meaningful relationship between radar observations and precipitation. In addition, the precipitation fields shown in Figures 6 and 10 appear to be dominated by artifacts rather than coherent precipitation structures. Taken together, these results do not provide evidence that the proposed method produces reliable precipitation estimates. Given these issues, I cannot recommend publication of the manuscript in its current form.
General Comments
Specific Comments