Water Resources Research Research Article · 62, e2025WR041529 Open Access · CC BY-NC

Depth-Aware Global Calibration of SM2RAIN-NWF Using Growing Neural Gas-Derived Hydroclimatic Clusters Across Heterogeneous Soils

Mohammad Saeedi1, Subin Kim2, Euiyoung Choi2, Hyunglok Kim2✉, Venkataraman Lakshmi1
1Department of Civil and Environmental Engineering, University of Virginia, Charlottesville, VA, USA · 2Department of Environment and Energy Engineering, Gwangju Institute of Science and Technology (GIST), Gwangju, South Korea
https://doi.org/10.1029/2025WR041529 · Received 2 Jul 2025 · Accepted 31 Mar 2026 · Special Collection: Advances in large-scale hydrological modeling and prediction under global change
The thesis in one line
Bottom-up rainfall retrieval normally needs a local rain gauge to calibrate every station. Cluster the world into 30 hydroclimatic zones from ERA5-Land, and a single set of cluster-average coefficients (Z̄, ā) reproduces fully calibrated skill — so SM2RAIN-NWF can run where no gauge exists.
Calibration data
0
reference rainfall needed at the target site
Best R gain (GNG)
+17.7%
cluster 21, 5–10 cm
K-means POD shift
≈0%
0.00–0.85% in most cases
01The gap the calibration phase is what keeps bottom-up retrieval out of data-sparse regions

Gauges are sparse and uneven

The rain gauge is still the primary instrument, but the network is sparse, concentrated in developed countries, and carries measurement error and poor spatial representation.

Top-down satellites have their own limits

IMERG and similar products face near-real-time aggregation and inter-calibration challenges across a heterogeneous satellite constellation, plus high operational cost.

Bottom-up works — but needs calibrating

SM2RAIN and SM2RAIN-NWF both require reference rainfall at spatial and temporal resolution comparable to the soil-moisture data, just to fix Z and a. That reliance constrains the method to gauged regions.

The circular problem

You need rainfall observations to build a model whose purpose is estimating rainfall where observations don't exist. This paper breaks that circle by transferring parameters from hydroclimatically analogous sites.

02Data ERA5-Land for the clustering space · ISMN for the parameters and the truth

ERA5-Land reanalysis — defines the zones

Period2015 – 2024, monthly Gridnative 0.1° → resampled to 0.25° by aggregated nearest-neighbour averaging Treatmenttemporally averaged to static spatial features; all variables standard-scaled Cold monthsretained, not excluded — snow depth and snowmelt keep the full annual hydroclimatic signature

Process groups screened

EvapotranspirationSoil moisturePrecipitationWindTemperatureRunoffSnowPressureVegetationLakeRadiationLat / Lon

Multicollinearity screening

Iterative Variance Inflation Factor screening, target VIF < 10, while keeping every process group represented. All variables passed except two, kept deliberately so their categories stayed in the feature space:

t2m · VIF 23.27str · VIF 13.7918 variables retained for UMAP and clustering.

A static soil-properties map was not an explicit feature; its influence enters through land-surface response variables (long-term volumetric soil water, runoff, evaporation) that integrate soil, geology and vegetation effects on the regional water balance.

ISMN in-situ network — supplies parameters and validation

Hourly soil moisture and precipitation records; stations assigned to clusters by geographic location. Station counts differ by measurement depth:

Global coverage is dense over North America and Europe, with smaller groupings in West Africa, East Africa, eastern China and south-eastern Australia (Fig. 1 of the paper).

Rain-event definition

For POD and FAR, daily precipitation is first converted to a binary rain / no-rain series with a 1 mm day⁻¹ threshold — any day above it (in situ or SM2RAIN-NWF) counts as a rain event.

03The self-calibration pipeline from reanalysis fields to a transferable parameter set
01
ERA5-Land features
monthly 2015–2024, resampled to 0.25°, time-averaged, standard-scaled
02
VIF screening
iterative removal to VIF < 10 → 18 variables spanning all process groups
03
UMAP embedding
non-linear manifold projection to a 3-D feature space preserving local and global structure
04
Clustering, k = 30
K-means · GNG · GMM · AC — four algorithms compared
05
Cluster-average Z̄, ā
ISMN stations assigned by location; per-cluster, per-depth means with bootstrap CIs
06
Apply to unseen sites
run SM2RAIN-NWF with the transferred parameters — no local calibration

Why k = 30

Two independent justifications converge: the Köppen–Geiger classification contains approximately 30 global climate subtypes, and GNG — which does not require k a priori — independently identified 30. The choice is grounded both climatologically and in the data.

Why four algorithms

K-means finds spherical clusters efficiently; GNG grows an adaptive topology capturing non-linear structure; GMM allows probabilistic, ellipsoidal assignment; AC reveals hierarchical nesting. Using all four tests whether the conclusion depends on the clustering method — it does not.

Two validation designs

LOOCV — for a cluster with N stations, recompute the cluster mean N times, each time excluding one station, then apply to the held-out station.
20% hold-out — split each cluster 80/20, derive (ā, Z̄) from the training subset with B = 10,000 bootstrap resamples for 95% CIs, apply to the untouched test set.

04The model being calibrated: SM2RAIN-NWF soil water balance inverted through an analytical net water flux

From water balance to rainfall

EQ 1 Z dθ(t)dt = P(t) − R(t) − ET(t) − G(t)
EQ 2 — classic SM2RAIN inversion P(t) = Zdθ(t)dt + a θ(t)b 1 − θ(t)c

Under the assumption that evapotranspiration is negligible during rainfall. A calibration procedure estimates all four parameters Z, a, b, c — and that procedure is exactly what this paper removes.

Zdepth-related soil parameter
θ(t)volumetric soil moisture
P, R, ET, Grainfall, runoff, evapotranspiration, drainage

What is actually calibrated here

Only Z and a. The fluctuation of c was found negligible — a stable value of 15.90 ± 0.005 across all stations — so c was held constant and the investigation restricted to the two sensitive parameters.

The NWF term — drainage done analytically

SM2RAIN-NWF replaces empirical drainage with the net water flux derived by inverting Warrick's (1975) solution of the linearized Richards equation. Reported to improve rainfall estimation accuracy by ~30% over SM2RAIN at basin (Lake Urmia) and national (Iran) scales.

EQ 3 — recursive net water flux FN = { ΘN(Z) e0.5Z−0.25TU(Z, ΔT)[N = 1] = F1 + ΘN(Z)e0.5Z−0.25TF1U(Z,T)U(Z, ΔT)[N = 2] = FN−1 + 1U(Z, ΔT){ ΘN(Z)e0.5Z−0.25TF1U(Z,T) − Σi=3N(Fi−1Fi−2) U(Z, [Ni+2]ΔT) }[N > 2]
EQ 4 — the estimator used P(t) = Z dθ(t)dt + ETpotθ(t) + a F(t) 1 − θ(t)c
Tdimensionless time
Udimensionless function from Warrick's solution
N, inumber of discrete time steps, and each step within the total N
05Which clustering wins? Table 1 — internal validation across the four algorithms

Three internal validation metrics, three separate scales

K-means GNG GMM AC
Silhouette ↑
cohesion vs separation
Davies–Bouldin ↓
average cluster similarity
Calinski–Harabasz ↑
between / within dispersion

K-means sweeps all three — highest silhouette (0.43), lowest Davies–Bouldin (0.82), highest Calinski–Harabasz (197,736). GNG is the clear runner-up (0.40 / 0.86 / 190,865) and beats GMM and AC on Calinski–Harabasz, so both were carried into the calibration analysis. GMM is weakest on every measure.

Table 1 — raw values

What the clusters look like

The 3-D UMAP embedding separates the ERA5-Land feature space into 30 well-defined groups for both K-means and GNG (Fig. 2a, 2c). Mapped back onto the globe (Fig. 2b, 2d), those groups form coherent geographic regions of shared hydroclimatic behaviour — which is precisely what makes a shared calibration coefficient defensible.

GNG produces clusters that differ from partition-based methods, reflecting subtle non-linear patterns — some clusters (e.g. 21) show distinctly wider parameter spreads, others (e.g. 2) much narrower, indicating more homogeneous soil or climatic properties.

A stated boundary condition

Only clusters with sufficient in-situ station coverage were retained. Clusters lacking adequate observational support were excluded to avoid unstable parameter estimation — so the framework needs a minimum of observations somewhere inside each zone, and is not yet directly transferable to regions with none.

06The deliverable: a lookup table of cluster coefficients Figs. 5b & 7b — mean Z̄ and ā per cluster, per depth
Algorithm
ā — mean drainage coefficient
grouped by soil-moisture depth interval
Z̄ — mean depth-related soil parameter
grouped by soil-moisture depth interval
≤ 5 cm 5–10 cm 10–20 cm 20–30 cm hover any bar for the exact value

This table is the product of the paper. Give a new, ungauged site its cluster and its sensor depth, read off (Z̄, ā), and SM2RAIN-NWF runs — no reference rainfall required. Z generally ranges higher in deeper layers (e.g. K-means cluster 23 at 10–20 cm, cluster 11 at 20–30 cm), whereas a may rise, fall, or turn bimodal depending on cluster composition.

Table view — all cluster coefficients, both algorithms
07How the coefficients behave with depth Figs. 3–4 — density against elevation, and histograms by cluster

≤ 5 cm

Z is broadly distributed at low elevations; a shows a sharp, highly concentrated peak at low values in the same elevation band — strongly constrained behaviour near the surface.

Diversity of calibration coefficients is highest here: a broad, near-uniform distribution across clusters.

5–10 cm

The distribution begins to concentrate at lower values, suggesting the parameter is becoming more constrained by underlying soil texture rather than diverse surface-level influences.

10–20 cm

The low-elevation distribution of Z becomes distinctly bimodal. Evapotranspiration, infiltration and soil layering likely play a greater combined role in setting the coefficients.

20–30 cm

The primary low-elevation mode consolidates toward the middle of the parameter range and a separate mode of lower Z appears at high elevations. The sharp surface peak in a has become broad and diffuse.

Many clusters show scattered patterns in both Z and a, possibly from significant changes in infiltration.

Why the spread grows downward

Wider violins and multi-peaked distributions (e.g. cluster 12 in some depth ranges, cluster 21 across multiple depths under GNG) indicate greater heterogeneity among stations within one cluster — the same zone can contain very different calibration behaviours. Narrow violins with tight boxes (cluster 4 at the surface; cluster 10 at shallow and mid depths; cluster 2 under GNG) mean more consistent station characteristics and more robust mean estimates.

The authors note a second, methodological contributor: Kim and Crow (2024) show that simplifying the water-balance physics and the limitations of soil-moisture data — noise and low sampling frequency — can themselves drive apparent variation in Z. The broadening seen in the deeper-layer clusters (notably 23 and 11) is probably partly this.

08Does removing calibration hurt? Tables 2, 3, A4, A5 — 20% independent station hold-out, % change relative to the fully calibrated baseline
Algorithm
Metric
cluster parameters are better than site calibration worse than site calibration

Full hold-out tables — coefficients with 95% CIs and all four metrics

GNG — the more compelling story

Some clusters show minor local degradation, but a significant number improve on the fully calibrated baseline. Most evident in correlation at 5–10 cm:

  • Cluster 2 — R +9.8%
  • Cluster 7 — R +15.3%
  • Cluster 21 — R +17.7%
  • Cluster 22 — R +8.5%
  • Cluster 21 at 20–30 cm — R +14.5%

Why improvement is possible: the cluster-average parameter is a more robust and generalizable estimator — it is less prone to overfitting the noise and idiosyncrasies of a single site's calibration record.

K-means — stability

Performance shifts are minimal across all depths and clusters. Changes in R are typically less than ±4%, and POD is almost perfectly preserved — 0.00% to 0.85% change in most cases.

Together with GNG's gains, this is the paper's central conclusion: cluster-based parameterization is a practical, robust pathway for scalable deployment in data-limited regions.

Consistent across metrics

Calibrated runs show systematically slightly lower RMSE and reduced bias, but the differences are not large. FAR and POD show no significant difference between the calibrated (green) and cluster-parameterized (red) cases in Figs. 6 and 8 — event detection survives intact.

GMM and AC — the same verdict, weaker margins

Both were carried through the same hold-out design (Tables A4, A5). They show the same qualitative result — cluster-derived coefficients hold up — but with larger negative excursions in R. Note the paper reports these tables without interpretive discussion; the reading below is drawn directly from their values.

  • GMM worst R change: −9.09% (cluster 2, 5–10 cm)
  • AC worst R change: −6.60% (cluster 5, ≤5 cm)
  • Both show consistently negative RMSE change — i.e. lower error than the calibrated baseline — up to −15.04% (GMM) and −13.26% (AC)

The paper's own method-independence claim rests on K-means and GNG, whose results it calls uniform across both frameworks. Tables A4 and A5 point the same way, with thinner margins.

09What to take away key points, limits, and where it goes next

The three key points

  • i. Cluster-based calibration enables SM2RAIN-NWF without additional site-specific rainfall data.
  • ii. ERA5-Land clustering defines the zones from which calibration coefficients are derived using in-situ data.
  • iii. Applying cluster-derived parameters to new sites maintains rainfall estimation accuracy.

The conceptual shift

From point-wise calibration to regionalized calibration: each region can be assigned a fixed — or narrowly distributed — parameter set for a and Z, instead of calibrating the model separately at every station.

Limits the authors state plainly

  • The framework needs a minimum level of observational support inside each cluster; clusters without it were excluded, and the method is not directly transferable to regions lacking in-situ data.
  • Clusters with wide distributions or overlapping confidence bounds need further refinement or additional explanatory variables.
  • Only Z and a are calibrated; c is held fixed at 15.90.
  • Two features (t2m, str) exceed the VIF < 10 target and were retained on representational grounds.
  • Soil properties enter only indirectly, through land-surface response variables.

Where it goes next

  • Transfer and pooling approaches to extend coverage to clusters with limited in-situ data.
  • Richer clustering variables: comprehensive land-use data, vegetation indices, human activity including irrigation, and subsurface hydrological dynamics.
  • Direct integration of high-resolution soil-property and geology data sets to refine the physical basis of the parameterization in complex terrain.
  • Higher spatial resolution: downscaled 1 km SMAP / AMSR2 (and 400 m) products, plus CYGNSS GNSS-R for sub-daily soil moisture.

Why it matters

It opens SM2RAIN-NWF to remote and data-limited areas, and supports scalable climate monitoring, reanalysis and prediction.