Shiwen Deng 1,2, Chengsheng Chen 1, Yunpeng Wang 1
(1. State Key Laboratory of Deep Earth Processes and Resources, Guangzhou Institute of Geochemistry, Chinese Academy of Sciences, Guangzhou, 510640, China.
2. University of Chinese Academy of Sciences, Beijing, 100049, China.)
Abstract: Natural hydrogen (H2) represents a promising low-carbon energy resource, yet interpreting its surface seepage remains challenging because surface signals integrate deep geological transport with complex near-surface environmental processes. In this study, we performed systematic multi-depth in situ gas measurements (300 sites) across the Sanshui Basin using a portable GA5000 analyzer. Nonparametric Mann–Whitney U tests and spatial autocorrelation analyses (Moran’s I = 0.763, Getis-Ord General G = 7.27, p < 0.001) revealed that while shallow measurements were fragmented, deep observations (>100 cm) exhibited strong basin-scale spatial clustering. Partial Mantel tests further confirmed significant correlations between deep H2 concentrations, fault influence weight (FIW, r = 0.053, p < 0.001), and topographic position index (TPI, r = -0.059, p < 0.001). Consequently, a co-kriging geostatistical framework incorporating FIW and TPI covariates successfully reconstructed the basin-scale deep hydrogen seepage field, resolving continuous high-flux seepage corridors aligned with primary NE-SW fault systems. To elucidate the mechanisms governing depth-dependent seepage variations, depth-resolved Pearson and partial correlation analyses were conducted across the soil profile, revealing a vertically structured filtering effect: upper soil layers (30-100 cm) are most strongly associated with silt content (partial correlation rpartial = -0.201 to -0.184), cation exchange capacity (CEC, rpartial = -0.199 to -0.231), and bulk density (BD, rpartial = -0.184 to -0.198), consistent with constraints on diffusive gas transport, whereas deeper layers (100-200 cm) are dominated by total nitrogen (TN, rpartial = -0.197), silt (rpartial = -0.208), and soil organic carbon (SOC, rpartial = -0.063), suggesting stronger biogeochemical attenuation. Finally, Random Forest (RF) and XGBoost models integrated with SHAP interpretability demonstrated that relying solely on Landsat 9 multispectral indices yielded limited predictive power (RF: R2 = 0.404; XGBoost: R2 = 0.407), whereas integrating multi-depth soil physicochemical properties substantially improved the predictive performance (RF: R2 = 0.705, RMSE = 104.6; XGBoost: R2 = 0.762, RMSE = 94.0). These findings indicate that near-surface soil properties play a key role in modulating the surface expression of natural hydrogen, highlighting that surface seepage represents a filtered signal of subsurface hydrogen transport and providing an important basis for natural hydrogen exploration.
Keywords: Natural hydrogen; Surface hydrogen seepage; Soil physicochemical properties; Vertical soil filtering; Spatial reconstruction; Sanshui Basin
Author Profile:
Shiwen Deng, Male, PhD student, mainly engaged in remote sensing applications and machine learning predictive modeling for natural hydrogen. E-mail: dengshiwen23@mails.ucas.ac.cn
