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

Abstract

Accurate rainfall information underpins land-surface water budgets, extreme-weather analyses, and climate-model evaluation. Yet in many regions, rain gauge networks are sparse, making conventional calibration of bottom up rainfall products difficult. To address this, we propose a self calibration framework that removes the need for a dedicated calibration phase. Our proposed approach systematically identifies bottom-up model parameters without relying on region-specific tuning by exploiting the K-means, Gaussian Mixture Model|Gaussian Mixture Models (GMM), Agglomerative Clustering (AC) and Growing Neural Gas (GNG) algorithms. To demonstrate its effectiveness, we apply this framework to soil moisture (SM) to RAIN by using Net Water Flux (SM2RAIN-NWF), a bottom-up rainfall estimation model that leverages SM variations to infer rainfall. This self-calibration strategy is particularly relevant, as it reduces dependence on traditional rain gauge data, making it well-suited for large or data-limited regions. In this study, we test this framework by comparing four clustering algorithms (K-means, GNG, GMM, and AC) against International Soil Moisture Network observations using hold-out and Leave-One-Out Cross-Validation approaches. This validation confirms the framework's robustness and the superiority of K-means and GNG. The K-means method provides high stability, with key performance metrics (Correlation Coefficient (R) and Probability of Detection) showing minimal change from the baseline. The GNG method demonstrates that cluster parameters can significantly outperform site-specific calibration, with correlation (R) gains exceeding +17% in key soil depths (5 < SM_Depth ≤ 10 cm). With no need for reference rainfall, the method is ideal for data sparse or ungauged regions and supports scalable climate monitoring, reanalysis, and prediction.