the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
| 12 Aug 2026
Increasing model parsimony without loss of process fidelity: a simplified LASCAM rainfall-runoff model
Abstract. Rainfall-runoff models are widely used to support water resource planning, yet many struggle under prolonged drought. During these periods, cumulative water deficits can strongly affect streamflow response by altering vegetation water use and contributing to multi-annual storage dynamics. The Large Scale Catchment Model (LASCAM) is well suited to representing these processes through its deep groundwater store and flux structure, but its large number of calibrated parameters increases computational cost and limits practical application. This study developed a simplified version of LASCAM by reducing the number of free parameters while retaining the structural features of the original model. Using the latest version of LASCAM (LASCAM22) as the starting point, we applied a staged simplification procedure that combined sensitivity analysis, parameter default-value testing, and stepwise parameter-fixing calibration. The reduced configuration was then compared with GR4J, Sacramento, and LASCAM22 in split-sample testing. Several low-sensitivity parameters were fixed with limited loss of performance, while the stepwise parameter-fixing calibration supported the 15-parameter configuration (LASCAM15) as the best compromise between dimensionality reduction and performance retention. In the split-sample test, LASCAM15 achieved stronger validation performance than LASCAM22 and outperformed GR4J and Sacramento during both calibration and validation. The resulting model is more parsimonious (i.e., fewer calibrated parameters), is more computationally efficient to calibrate, and provides more robust simulations for catchment modelling applications.
- Preprint
(1845 KB) - Metadata XML
-
Supplement
(4811 KB) - BibTeX
- EndNote
Status: open (until 07 Oct 2026)
- RC1: 'Comment on egusphere-2026-4407', Anonymous Referee #1, 22 Sep 2026 reply
-
RC2: 'Comment on egusphere-2026-4407', Anonymous Referee #2, 25 Sep 2026
reply
This is a good manuscript, and the authors may address the following comments:
1. In my opinion, LASCAM15 is mainly a parameter-reduction exercise, rather than a fundamentally new hydrological modelling framework. Accordingly some of the statements in abstract, introduction and conclusions need to be toned down.
2. The key validation is a single chronological split at 2011, and the authors themselves acknowledge that all models perform substantially worse in the drier validation period. However, the authors have not discussed a possible solution.
3. Overestimation of stremflow by approx 20% is not aligned with the claim of robust model.
4. The case study is South-West Western Australia, but a model may be claimed robust when it is applicable to other regions.
5. Low sensitivity is not equivalent to physical irrelevance or poor identifiability, and somehow this is not discussed in details.
Citation: https://doi.org/10.5194/egusphere-2026-4407-RC2 -
RC3: 'Comment on egusphere-2026-4407', Anonymous Referee #3, 27 Sep 2026
reply
General comment
The manuscript proposes a systematic procedure for reducing the number of freely calibrated parameters in LASCAM through parameter sensitivity analysis, parameter default-value testing, and stepwise parameter-fixing calibration. The extensive multi-catchment experiments provide a potentially useful basis for evaluating parameter reduction in LASCAM.
However, the contribution should be more precisely defined. The original model structure and process remain unchanged; selected parameters are fixed at default values while the remaining parameters are recalibrated. Therefore, the study primarily develops a parameter-reduction/parameter-fixing method rather than a structurally simplified version of LASCAM. This distinction should be reflected consistently throughout the manuscript.
More importantly, several methodological assumptions underlying the proposed parameter-reduction framework require stronger justification. In particular, the necessity and added value of the pairwise experiments in Step 2, the applicability of the selected default values across catchments, and the dependence of the final reduced configuration on the sequential parameter-fixing order remain unclear. These issues are central because they determine whether the proposed reduced parameter set can reasonably be considered transferable beyond the catchments used to derive it. Furthermore, because the manuscript makes process-related claims concerning evapotranspiration, vegetation water use, and long-term storage, evaluation based almost exclusively on streamflow is insufficient to support these conclusions.
Overall, I consider the parameter-reduction framework potentially useful, but the manuscript requires substantial revision to better define its methodological contribution, justify the parameter-fixing procedure, evaluate the reliability and transferability of the selected parameters, and strengthen the validation of the reduced LASCAM configuration.
Detailed comments:
1 Introduction
Lines 40–43: The detailed description of individual LASCAM storage components is more appropriate for the model-description section. The Introduction should provide only the information required to motivate the study, while the detailed description should be moved to Sect. 3.
Lines 60–63: The concluding statement does not follow directly from the preceding discussion. The paragraph mainly discusses computational cost, parameter complexity, and the limited benefit of increasing model complexity, whereas the conclusion shifts to parameter non-uniqueness and uncertainty outside the calibration period. Please revise the paragraph so that the motivation for reducing the number of calibrated parameters is developed consistently.
2 Study area and data
Lines 92–95: Please explain why these meteorological datasets were selected. For example, were they chosen because of their spatial and temporal coverage, data quality, or suitability for Australian catchments? The spatial and temporal resolutions of the forcing and streamflow datasets should also be provided here.
3 LASCAM structure
Lines 114–119: This description substantially overlaps with material in the Introduction. I suggest retaining the detailed model description here and shortening the corresponding material in the Introduction.
4 Methods
4.1 Model simplification method
4.1.1 Overview
Lines 135–145: Clarify the terminology used for “model simplification”. The procedure does not appear to remove model processes or modify the governing equations; rather, it reduces the dimensionality of the calibration problem by prescribing default values for selected parameters. The manuscript should make this distinction explicit and consistently describe the proposed framework as parameter-reduction where appropriate.
4.1.2 Step 1: Sensitivity analysis
Please provide a clear table or cross-reference showing which parameters were identified as candidate fixed parameters after the sensitivity analysis.
4.1.3 Performance metrics and objective functions
The description of standard performance metrics and their equations could be shortened, with appropriate references provided where possible.
4.1.4 Step 2: Default-value selection
Lines 202–211: The necessity and added value of Step 2 are unclear. As I understand the procedure, Step 2 uses pairwise parameter experiments to determine whether a candidate fixed parameter and its default value remain acceptable when another parameter is allowed to vary. However, Step 3 subsequently fixes parameters sequentially and evaluates their effects through catchment-specific recalibration. It is therefore unclear why the pairwise experiments in Step 2 are required rather than using a simpler approach to determine candidate default values followed directly by the evaluation in Step 3.
If Step 2 does not independently determine the final parameter configuration, is it essentially a prior screening step for Step 3? Could similar candidate default values be obtained more directly from the calibrated parameter distributions across catchments and then evaluated through the stepwise recalibration in Step 3? Please clarify what additional information Step 2 provides and how it improves the reliability of the final parameter-fixing procedure.
4.1.5 Step 3: Stepwise parameter-fixing calibration
Lines 243–245 and Fig. 5: Clearly explain what is meant by the"21-parameter model”, “20-parameter model”. All numerical labels and symbols in Fig. 5 should also be explained in the caption.
5 Results
5.2 Parameter pairings and default values
Lines 313–315: The determination of universal default values from the pairwise experiments requires further justification.
For example, Fig. S6 shows substantial differences in the range of deltaf among catchments, with catchment 609002 exhibiting a markedly different range from several other catchments. Similarly, Fig. S2 shows substantial inter-catchment differences in the acceptable range of betaa.
These cases raise an important question regarding the criterion used to define a universal default value. The authors should therefore specify what quantitative criterion defines sufficient overlap among catchments; and how much deterioration in individual-catchment performance is considered acceptable when adopting a common fixed value.
5.3 Sequential parameter-fixing calibration
Lines 355–356: The criterion used to identify parameters that are considered unsuitable for fixing is unclear and should be explicitly described. The finding that the 15-free-parameter configuration provides the best performance does not necessarily demonstrate that the parameters are intrinsically unsuitable for fixing. A parameter that causes performance deterioration under one combination of fixed and free parameters may still be suitable for fixing under another configuration.
The authors should therefore clarify how parameter fixability is determined at each stage and assess whether the selected 15-parameter configuration is sensitive to the adopted fixing order.
5.5 Split-sample test
Lines 382–395: The primary comparison should be between LASCAM15 and the original LASCAM22, because this directly tests whether reducing the number of freely calibrated parameters preserves model performance. Better performance of LASCAM15 relative to GR4J and Sacramento does not demonstrate the benefit of the proposed parameter-reduction procedure, because LASCAM22 itself may already outperform these models.
I therefore suggest emphasizing the performance differences between LASCAM15 and LASCAM22 across calibration and validation periods. GR4J and Sacramento may remain as secondary benchmarks.
6 Discussion
6.1 Parameter sensitivity and implications for LASCAM structure
Lines 403–415: The discussion should avoid interpreting parameter reduction as structural model simplification. Since the model equations and represented processes remain unchanged, the demonstrated benefit is primarily a lower-dimensional calibration problem, rather than a structurally simpler hydrological model. Please revise the terminology and interpretation accordingly.
6.2 Retention of vegetation water use and long-term storage processes in LASCAM15
Lines 446–451: Hydrological processes are interdependent, and parameter fixing may alter ET, soil moisture, and storage dynamics even when their governing equations remain unchanged. If the authors wish to make process-level claims, LASCAM15 and LASCAM22 should also be evaluated using ET and soil moisture, preferably against independent observations where available.
Lines 459–461 and Figure 11a: The interpretation of Figure 11a should be revised. A more informative analysis would directly quantify and discuss the differences in simulated storage between LASCAM15 and LASCAM22. For example, Figure 11a appears to indicate systematically higher storage in LASCAM15 than in LASCAM22. The authors should discuss why fixing parameters leads to this difference, whether the magnitude and temporal variability remain physically plausible, and what this implies for the internal water balance of the reduced-parameter configuration.
Figure 11b: The comparison with groundwater observations also requires reconsideration. The observational data shown in Figure 11b represent groundwater level (m), whereas the model variable discussed in Figure 11a represents groundwater/deep storage. These are physically different quantities and cannot be directly compared solely on the basis of similar temporal trends.
The authors should simulate or derive an equivalent groundwater-level or water-table-depth variable from both LASCAM15 and LASCAM22 and compare it directly with the observations.
6.3 Extrapolating hydrological models to future conditions
Lines 491–493: The discussion of future applications should go beyond recommending additional calibration variables. The authors should also critically examine whether the proposed parameter-fixing procedure necessarily improves model performance under conditions outside the calibration domain.
Parameter compensation identified under the historical calibration climate does not necessarily remain valid under future climatic conditions.
The authors should therefore discuss whether fixing parameters based on historical compensation relationships could preserve streamflow performance while altering internal process responses under non-stationary conditions.
6.4 Broader implications and recommendations
Lines 516–517: The authors should distinguish clearly between the potential generality of the parameter-reduction methodology and the transferability of the specific parameter values obtained in this study. Independent evaluation under different hydroclimatic conditions and forcing datasets would be required before the latter can be considered broadly transferable.
7 Conclusions
The main contribution should be described as a parameter-reduction or parameter-fixing framework for LASCAM, rather than the development of a structurally simplified LASCAM model.
The Conclusions should also report quantitative changes in key performance metrics between LASCAM22 and LASCAM15 (e.g., NSE-Bias and PBIAS), rather than relying on qualitative statements that model performance was maintained or improved.
Citation: https://doi.org/10.5194/egusphere-2026-4407-RC3
Data sets
SILO climate database Queensland government https://www.longpaddock.qld.gov.au/silo/point-data
Water Information Reporting database Department of Water and Environmental Regulation (DWER) http://wir.water.wa.gov.au
CAMELS-AUS v2 Keirnan Fowler et al. https://doi.org/10.5281/zenodo.14289037
Model code and software
LASCAM_simplification Ziqi Zhang https://github.com/potatohey/LASCAM_simplification/tree/v1.0
Viewed
| HTML | XML | Total | Supplement | BibTeX | EndNote | |
|---|---|---|---|---|---|---|
| 140 | 89 | 26 | 255 | 62 | 18 | 19 |
- HTML: 140
- PDF: 89
- XML: 26
- Total: 255
- Supplement: 62
- BibTeX: 18
- EndNote: 19
Viewed (geographical distribution)
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
General comments
The authors fix seven of the 22 LASCAM22 parameters at shared default values. The defaults are chosen through Sobol screening and pairwise behavioural sampling on seven catchments, and the effect of fixing them is then tested by stepwise recalibration across 76 catchments in south-west Western Australia. Making an operational model easier to calibrate without changing its equations is a sensible goal, the regional sample is large, and the code is openly available.
My main reservation is that the strongest claims go further than the current analysis supports. As far as I can tell, the defaults and the choice of 15 parameters were informed by data from the evaluation period. The comparisons are not supported by paired statistics, process fidelity is shown for a single catchment only, and the efficiency gain is not measured. I think these points can be addressed, but several of them need new analysis rather than rewording.
Major comments
If I understand Sect. 4.1.5 correctly, all candidate configurations were calibrated on the full record (L247), and these calibrations also supplied the reference sets for the pairwise analysis (L208 to 210). The post-2011 data used for validation in Sect. 4.2 would then already have shaped both the defaults and the choice of LASCAM15. The note at L274 does not quite resolve this, because the influence enters earlier in the workflow. The period used for the Sobol screening is also not stated. I would ask the authors to confine every data-dependent decision (screening, reference calibration, default selection and configuration choice) to the pre-2011 period, fix those choices, and only then evaluate on the later period. If the selected configuration changes as a result, that is worth reporting in its own right. If this separation was in fact already in place, a clear statement of which data were used at each step would be enough.
The acceptance rule in L257 ("little or no decline") is not defined, and the median decline below 0.1 reported at L338 describes the outcome rather than showing that it is acceptable for planning use. I would like to see catchment-level paired differences against LASCAM22, the share of catchments that deteriorate noticeably, and an uncertainty estimate that allows for nested or neighbouring catchments. It would also help to state an acceptable loss in advance for the revised analysis. On Fig. 10, LASCAM15 has the higher validation median only for Split-Trotter. Its NSE-Bias median looks slightly lower than that of LASCAM22 (roughly 0.25 against 0.35), and the PBIAS medians are similar. In calibration, LASCAM15 and GR4J are close on NSE-Bias, and Sacramento has the median PBIAS nearest zero. The wording at L380 and in the abstract should therefore be adjusted metric by metric.
The rule for accepting a default shifts between "most" catchments (L220), "any" catchment (L221 to 222) and "all" catchments (Fig. 4 caption), and Sect. 5.2 relaxes it after the results are in. Please state one rule, explain how exceptions and the choice of partner parameter were handled, and give the number of qualifying catchments in Table 3. The 80% criterion (L216) also behaves oddly when the reference score is negative, since 0.8 times a negative optimum lies above the optimum. Because each grid is a slice around a single reference solution, I would also like to know whether the defaults hold up when other near-optimal sets from the 30 restarts are used, and in a leave-one-catchment-out test. The seven primary catchments were chosen for their long records and low disturbance, so testing the defaults on catchments not used to derive them seems important.
LASCAM15 is exactly the point at which all parameters with identified defaults have been fixed, and the 14- to 12-parameter runs rely on provisional values (L333 to 335). I am not sure the decline after 15 reflects a genuine need for those degrees of freedom rather than poorly chosen provisional values. Could the authors explain where these values came from? The gammab examples in Fig. 7c and 7d seem to sit near 0.1 and near 0.35 to 0.47, which does not obviously support 0.2. A comparison at the same dimension against a simple alternative, for instance the medians of the LASCAM22 calibrated values, would show whether the pairwise procedure adds anything. I also noticed that the NSE-Bias spread widens mainly when alphac is fixed (the 19-parameter step in Fig. 8b), so a short test that releases alphac would be informative. Unless further evidence emerges, I would present 15 as a reasonable choice under the tested settings rather than as an optimum.
Please state whether the roughly 150,000 runs (L163) are per catchment, give the base sample size and sampling scales, and add bootstrap intervals or a convergence check. Two details matter here. First, if amn (0 to 500) and amx (80 to 1000) were sampled independently, about one fifth of the samples have amn larger than amx, which makes the denominator in Eq. B3 negative. Was the ordering enforced? Second, if betaa was sampled linearly over 0.01 to 10000, nearly all sampled values exceed 10. More generally, a low total-order index across the full prior does not guarantee that a parameter is unimportant near the optimum. Fig. 7c is a case in point: at Scott River the score falls sharply once gammab exceeds about 0.15, even though gammab ranks 19th on average. I would therefore be cautious about equating low sensitivity with limited identifiability (L416).
Fig. 11 covers one catchment, the bore record has a long gap, and a decline in storage under declining rainfall would be expected from many models with a slow store. Agreement with LASCAM22 is also not the same as agreement with observations.
Please quantify, for each catchment, the change in precipitation and PET between the two periods. The map in Fig. 1 uses different periods, and the evaluation period is described as both "typically drier" (L271) and "significantly drier" (L394). A few additional splits would show whether the results depend on 2011 as the break point. Relating the magnitude of drying to the difference between LASCAM15 and LASCAM22 could be an informative exploratory analysis, though not a causal test. The over-parameterisation interpretation (L532 to 533) first appears in the Conclusions. Evaluating a few intermediate configurations in the split-sample test would help here; an optimum at intermediate complexity would be quite consistent with that interpretation.
The Introduction mixes calibration cost, identifiability and transfer under change without saying which of these the study actually tests, and t
he question of process retention appears only in Sect. 6.2. Stating the research questions at the end of the Introduction would sharpen the contribution, as would placing the work alongside earlier studies of sensitivity-guided parameter reduction (e.g., Tang et al., 2007; van Werkhoven et al., 2009; Cuntz et al., 2015). I would also move the groundwater analysis into the Methods and Results, add a short limitations paragraph, and let the known limitations, including the roughly 20% median overestimation after 2011 (L478), temper the abstract and conclusions. Phrases such as "demonstrating the stronger robustness" (L383) and "confirm that the model is quite flexible" (L484) go beyond what the analysis shows.
Minor comments
The units in Tables 1, 4 and A1 need checking. amn, amx, bmx and fmx scale storage and should be in mm rather than mm/day, and alphag, fs, alphaf, betab and td are not dimensionless. Eq. B6 also appears to cap a flux with the dimensionless parameter mu. The constants in Eqs. B5 and B17 should be documented.
Eq. 3 has an extra summation in the denominator, B in Eqs. 1 and 4 is not defined explicitly, and the range of the bounded Trotter score should be (-1, 1].
A table of the 76 catchments would help, listing ID, area, data source, record period, valid calibration and evaluation years, membership of the development set, and known human impacts. Please also name the SILO PET product and state whether the forcing is taken at a point or averaged over the catchment.
Fig. 6 has no readable radial scale. Figs. 8 and 10 should define the whiskers and show outliers. Fig. 8a seems to have a lower whisker near 0.25 at 22 parameters, while Fig. 9 shows catchments with negative scores. Difference maps would complement Fig. 9, and the Fig. 1 legend class "-50 - 25" should read "-50 to -25".
In Fig. 7 the scanned axis is labelled "(fixed)", which is a little confusing, and the colour ranges hide differences among good solutions. Table 3 should define "/", list the provisional values and overlap counts, and flag that bmx sits at its upper bound. Footnote a of Table 4 seems at odds with keeping those bounds in the recommended configuration.
For Fig. 11, please define "m bTOC" and its sign convention, state which calibration is shown and whether the warm-up period is included, avoid lines drawn across long data gaps, and check L463 to 464 against the data plotted around 2010.
The title should include a version number, in line with GMD practice. The eWater Source plugin mentioned at L66 to 67 is not covered by the Code availability statement.
References: please update Burns et al. to the 2026 HESS paper and reconcile Trotter et al. 2021 (L56) with the 2022 entry in the reference list. The "Gilles, Arsenault and Brissette" entry appears to be Essou et al. (2016), cited at L277. LASCAM22 is referenced to a Zenodo dataset (L104), so its description should be complete within this paper.
The captions of Figs. 4, 6 and 7 state conclusions; they would work better as descriptions of what is shown.
A few language points: "not all planned tested could even be complected" (L57 to 58), "each of them have" (L81), "alphaq" (L410), "few flux towers nor groundwater wells" (L497) and "it is pleasing that" (L384). The spelling of terms (Split-Trotter, NSE-Bias, Sacramento, LASCAM15) varies, and some passages are repeated (L7 and L23, L43 to 47 and L114 to 119, L92 to 95 and L649 to 652).