the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 26 Feb 2026
Shifts in the Dominance of Climatic, Anthropogenic, and Landscape Drivers Explain the Spatial Variation in the Budyko-type Equation Parameter
Abstract. Hydroclimatic transition zones are critical hotspots of global environmental change, yet the spatial heterogeneity of their hydrological functioning remains poorly understood because of the complex interplay of natural and anthropogenic factors. In this study, we propose a machine learning-driven diagnostic framework to better understand the spatially divergent drivers of the Budyko parameter (ω) across 12 representative catchments in the semi-arid to semi-humid transition zone. By integrating principal component analysis with hierarchical clustering, we objectively identified three distinct hydrological functional zones. Four machine learning algorithms (XGBoost, RF, ANN, and SVM) were subsequently systematically benchmarked for each zone to select the optimal model, and Shapley Additive exPlanations (SHAP) analysis was performed to quantify the driving mechanisms. The results reveal a fundamental spatial shift in the dominant drivers of ω: C1 is dominated by climatic factors (53.52 %), C3 is dominated by anthropogenic factors (68.73 %), and C2 is jointly driven by climatic (37.24 %), anthropogenic (31.85 %), and landscape (30.91 %) factors. Specifically, the primary drivers for ω are temperature (T, 32.42 %) in C1, leaf area index (LAI, 24.59 %) in C2, and GDP (25.09 %) in C3. The critical thresholds for shifting the directional contribution of these factors were 8.17 °C, 1.16, and 25.8×104 USD, respectively. Furthermore, the directional impacts of climatic, anthropogenic, and landscape drivers vary significantly across zones, with pairwise interactions exhibiting distinct patterns of synergy and trade-offs. This study demonstrates that local landscape characteristics and human activity patterns can override macroclimatic controls, providing support for spatially differentiated water resource management in climatic transition zones.
- Preprint
(2214 KB) - Metadata XML
- BibTeX
- EndNote
Status: final response (author comments only)
-
RC1: 'Comment on egusphere-2026-969', Cyril Thébault, 13 May 2026
-
AC1: 'Reply on RC1', Bingbing Ding, 11 Jul 2026
We sincerely thank Reviewer #1 for the constructive and detailed comments on our manuscript. We have carefully considered all comments and have substantially revised the manuscript accordingly. The revisions mainly include improving the methodological transparency and reproducibility, clarifying the PCA-based functional zoning procedure, providing more details on the model training and validation strategy, adding hyperparameter information and supplementary materials, and clarifying the implementation and interpretation of SHAP analysis.
We also revised several figures and captions to improve readability and avoid possible misinterpretation. In particular, we added a schematic figure to explain the Budyko parameter, revised the methodological framework figure, replaced or redesigned several performance and SHAP-related figures, and clarified the meaning of model outputs, SHAP contributions, transition points, and directional alignment patterns. Throughout the revised manuscript, we also moderated statements that could imply direct physical causality or deterministic mechanisms, and we added more explicit discussion of limitations related to the small number of catchments, predictor collinearity, and uncertainty in SHAP-based interpretations.
A detailed point-by-point response to all reviewer comments is provided in the attached supplement. Please note that, following the journal instruction, the revised manuscript itself is not submitted here as a supplement and will be provided through the formal revision system when requested.
-
AC1: 'Reply on RC1', Bingbing Ding, 11 Jul 2026
-
RC2: 'Comment on egusphere-2026-969', Anonymous Referee #2, 03 Aug 2026
The framework presented by the authors has the potential of a valuable contribution. However, I see a lot of methodological issues, ranging from missing clarity, clear justification, appropriatness & motivation up to unsopported later claims. The methodology ultimately translates into strong hydrological conclusions & claims that in my opinion are barely supported. More precisely, the "critical" thresholds (the papers headline) reported seem often contradictory to the hydrological/physical understanding and the evidence is not very high.
My recommendation is therefore rejection with the potential of resubmission of a completely revised version.
Here are my detailled comments, a PDF version with further statements is additionally attached:
1. Potential train/test data leakage that likely blows up accuracy (Table 2, Figs. 5–6)
ω is derived from 11-year moving windows, so consecutive samples share ~10/11 years of raw data. The 70/30 split (Sec. 2.6) is random, not chronological if I understood the methodology correctly. So "test" rows sit next to near-duplicate training rows I assume? This likely explains R² up to 0.995 which does not indicate a well gernalized model but a highly overfitted model... The paper's own Limitations mention leave-one-basin-out (LOBO) results showing instability which is essential. This belongs in the main results, not a closing statement. In any case the sampling seems to cause issues in data independence.2. Sample size is very, very small with 12 catchments and the later clusters of 3/4/5
If I understood the methodology properly the authors end up with very few clusters based on a very limited number of basins (N=12), however, come up with very strong and confident statements on hydrological threshold behavior in the semi-arid/semi-humid transition zone. I think the first thing worth to flag is, either the authros need a complete reframing and treat the whole paper more as a case study in China or increase the number of basins much more to allow for a statistically sound version of the paper that also allows to draw generalized conclusions.
Besides, reporting SHAP thresholds for 3 different variables (e.g., T=8.17°C, LAI=1.16, GDP=25.8×10⁴) implies precision the data in my opinion the data can't support. At least this needs kind of leave-one-catchment-out sensitivity checks and authors should probably report ranges, not point estimates, especially for C2 (N=3 basins).3. "Transition zone" character not fully demonstrated
The study area section was a bit surprising to me, as no per-catchment climatology table or cited zoning method is given before clustering (Sec. 2.1). I indeed was a bit puzzled that the underlying climate classification map which is overlaid by the basins is not even referenced!? Is it KKöppen-Geiger, a Chinese product, or an own independent illustration from the authors? Using the later cluster-mean AI values (C1=0.51, C2=0.58), both fall in "dry sub-humid," not clearly "semi-humid." Some catchments (e.g., "l") do not look like transitional cases but are located fully and clearly in the corresponding regions, so the basin choice is not really justified. I think this risks even circularity: the authors put as an objective the functional zoning, however, the classification basically reveals what was already there by design (given the basin choice) so I am not sure whether the clustering was even needed, revealed something new or was even statistically sound (given the fact that the PC did not really distinguish well as we see climatic but also anthropogenic features).4. ω is analytically derived, not observed which can affect both: target and predictors
ω is back-calculated from P/R/E0 via Fu's equation, so any unmodeled process (e.g., reservoir regulation, absent from the 13 features) is absorbed directly into the target itself, not just omitted as a predictor, right? Additionally, P and AI are used both to derive ω and as ML predictors, so part of the reported "driver importance" (Sec. 3.4) partly re-derives Fu's equation rather than finding an independent empirical pattern or am I missing something?5. Reservoir/dam regulation might be missing with huge implications?
GDP density correlates with dam/hydropower infrastructure, which the model can't distinguish from the proposed "ecological investment" mechanism (Sec. 4.2). Currently only flagged as a one-line caveat tied to groundwater in C2 (line ~446), not addressed for C3's GDP interpretation.6. Section 2.5 (PCA/clustering) lacks transparency
- n, p, and matrix construction (per-catchment vs. per-window) aren't stated explicitly.
- Silhouette scores are assumed with "k=3 optimal" without reporting any scores across cluster sizes. Generally there is not much stated on the method, not sure if anyone unfamiliar with the model has a chance to know what is going on.
- The necessity and added value for clustering is completely missing and in best only appears later in the Discussion (Sec. 4.1, via Fig. 10), not in Methods where it's needed.
- PC retention rule not clear to me. Section 2.5 never says how many PCs were kept or by what criterion (e.g., eigenvalue >1, cumulative variance threshold). The 85.1% cumulative variance figure for PC1+PC2 comes later and one might infer there was a threshold.
- PCA (Fig. 4a) shows climate-wetness and development-intensity onto the same PC1 axis which might imply the design can't cleanly separate "wet" from "developed," which complicates attribution?
7. Figs. 5 and 6
Fig. 5's violins are per-catchment aggregate scores are only 3–5 points per cluster, right? If so I guess the violin plots do not make sense, however, it is not clear (barely anywhere) what the N of the individual plots is and why. Fig 6 on the other hand is even unclearer to me, what is shown there, test results per classification zone? Meaning w values from the 30% testing data from the 1985-2015 period?8. Reported thresholds do likely not hold
With only 3–5 catchments per cluster, these curve shapes cannot be distinguished from artifacts driven by one or two catchments. It does not seems statistically sound and the drawn conclusions probably do not hold and are somethign for large-sample analysis. E.g. for several features, the paper's SHAP curves show a shape that contradict standard hydrological reasoning which is surprisingly never addressed and properly discussed:- T: In an energy-limited regime (cool, high-slope catchments, e.g., C1's mean T is 7.89°C), I would argue every additional degree should gradually increase ET and thus ω, since energy is the binding constraint!? Instead, the paper reports SHAP values that are negative below 8.17°C and only turn positive above it, which is a sharp surprising switch, not something the theory would assume!? No explanation is given for why warming would decrease ω at low temperatures.
- LAI: Standard assumption would be more LAI → more transpiration → more ET → higher ω. However, in the paper, LAI > 1.16. leads to the opposite: higher LAI is associated with lower ω. The only explanation offered is a guess ("likely" soil saturation or agricultural drainage) with no supporting evidence.
- GDP: the low-GDP segment (negative SHAP) has no explanation at all only the positive/high-GDP segment is discussed.
- AI in C2 is the one exception: both segments of its curve are explained by a standard water-limited-to-energy-limited transition.
9. Causal language
Sec. 3.2 (cluster description) already uses interpretive phrases ("physical basis," "fragile ecosystem," "shaped by... human development") before any mechanism analysis (Sec. 3.4/4.2) has been done.Besides, further points that puzzled me is for example:
- the population density of 1.1*10^7 per km2 -> did the authors actually check the numbers? 11 million person per km2 seems more than suspicious...
- ω framing as "landscape-only" might be too strong, given many studies that demonstrated the impacts of intraannual climatic conditions
- No code availability statement given the complexity of the ML/SHAP workflow!?
-
AC2: 'Reply on RC2', Bingbing Ding, 25 Aug 2026
We sincerely thank you for the careful and technically rigorous assessment of our manuscript, including the detailed annotations in the accompanying PDF. We have carefully considered all of the concerns raised. A complete point-by-point response is provided in the attached supplement.
The review has helped us define the contribution of the study more precisely. The central scientific question is whether the relative associations of climatic, anthropogenic, and landscape factors with the effective Budyko parameter vary among catchments across the hydroclimatic transect of northern China. The intended contribution is therefore a regional multi-catchment comparison and a reproducible analytical framework, rather than the identification of universal thresholds or independent causal mechanisms.
We recognize that the reliability of this comparison depends particularly on the dependence among overlapping moving windows, the effective number of independent catchments, and the model-derived nature of ω. We have therefore developed a focused and feasible revision plan. The primary model evaluation will pool the data from all 12 catchments and use leave-one-catchment-out validation, complemented by temporally blocked validation that prevents overlapping years from entering both the training and test sets. Functional zones will be used for stratified interpretation, while separate within-zone models will be treated only as sensitivity analyses. This design preserves the regional comparison without relying on primary models fitted to only three to five catchments within individual zones.
Regarding population density, our re-examination confirmed that the underlying gridded population totals were not affected. The implausible values resulted from an incorrect denominator in the final density conversion for one group. After applying the correct basin-area denominator, our preliminary reanalysis retained the same three-zone membership. The population-density error therefore does not appear to determine the functional zoning, although the relevant analyses will be rerun and verified using the corrected input before any final conclusions are drawn.
The remaining methodological and interpretive concerns will be addressed through clearer reporting of the data and sample structure, robustness and sensitivity analyses, more appropriately calibrated hydrological interpretations, and improved computational reproducibility. Only patterns that remain stable under the corrected data and leakage-resistant validation will be retained. Detailed responses and the corresponding revision strategy for every comment are provided in the attached supplement.
We again thank the referee for these constructive comments, which have helped us improve the rigor, transparency, and scientific focus of the study.
Viewed
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 723 | 384 | 86 | 1,193 | 59 | 62 |
- HTML: 723
- PDF: 384
- XML: 86
- Total: 1,193
- BibTeX: 59
- EndNote: 62
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
This manuscript presents a machine learning-based framework to investigate the spatially heterogeneous drivers of the Budyko parameter across hydroclimatic transition zones in the Yellow River Basin, combining clustering techniques, ML models, and SHAP interpretation methods. The manuscript is generally well structured, the figures are of good quality, and the study addresses a relevant topic for the HESS community.
However, several aspects require substantial clarification and revision before the manuscript can be considered for publication:
Additional detailed comments and annotations are provided directly in the annotated PDF attached to this review.