Soil moisture is a key factor in hydrological, biological, and chemical processes, and plays a critical role in maintaining ecosystem balance. To generate high-resolution soil moisture maps at regional scales, researchers primarily employed in-situ observation-based spatial interpolation and remote sensing-based downscaling methods. However, direct comparisons between these methods are scarce. Additionally, remote sensing techniques are limited to the topsoil layer, and in-situ observations often have large depth intervals, thereby constraining the vertical resolution of subsurface soil moisture mapping. To address these challenges, we utilized an equal-area spline depth function combined with machine learning to map high spatial-vertical-resolution daily soil moisture across the Qinghai-Tibet Plateau. The performance of spatial interpolation and downscaling methods in mapping surface soil moisture at 0–5 cm depth were also compared. The results revealed that both spatial interpolation and downscaling methods produced unbiased predictions. However, prediction accuracy was lower in the peripheral subareas of the study area which had lower sampling density. Maps generated through the spatial interpolation method better captured detailed environmental covariates, whereas those obtained with downscaling methods were smoother. The fitting of depth functions introduced only small errors, but caution is still needed when predicting at unobserved depths. For subsurface soil moisture mapping using depth functions combined with spatial interpolation, validation results at two depth intervals showed improvements over surface predictions, with a root mean squared error (RMSE) reduced by 6.45 % to 17.2 % and unbiased RMSE by 5.95 % to 19.04 %. Furthermore, the analysis of variable importance highlighted the critical role of time-varying covariates. Future research should focus on optimizing depth functions and combining data-driven with knowledge-driven approaches. This study serves as a reference for mapping soil moisture with fine spatial-vertical-resolution in large-scale study areas.

中文翻译：

---

使用机器学习和深度函数在 3D 空间和时间中绘制高分辨率土壤水分图


土壤水分是水文、生物和化学过程中的关键因素，在维持生态系统平衡方面起着关键作用。为了在区域尺度上生成高分辨率的土壤水分图，研究人员主要采用了基于原位观测的空间插值和基于遥感的降尺度方法。然而，这些方法之间的直接比较很少。此外，遥感技术仅限于表土层，原位观测通常具有较大的深度间隔，从而限制了地下土壤水分测绘的垂直分辨率。为了应对这些挑战，我们利用等积样条深度函数与机器学习相结合，绘制了青藏高原的高空间垂直分辨率每日土壤水分图。还比较了空间插值和缩小方法在绘制 0-5 cm 深度表层土壤水分方面的性能。结果表明，空间插值和缩小方法都产生了无偏预测。然而，在采样密度较低的研究区外围子区域中，预测精度较低。通过空间插值方法生成的地图更好地捕获了详细的环境协变量，而使用缩小方法获得的地图则更平滑。深度函数的拟合只引入了很小的误差，但在未观察到的深度进行预测时仍然需要谨慎。对于使用深度函数结合空间插值的地下土壤水分制图，两个深度间隔的验证结果显示比地表预测有所改善，均方根误差 （RMSE） 降低了 6.45 % 至 17.2 %，无偏 RMSE 降低了 5.95 % 至 19.04 %。 此外，变量重要性的分析强调了时变协变量的关键作用。未来的研究应侧重于优化深度函数以及将数据驱动与知识驱动方法相结合。本研究为在大尺度研究区以精细的空间垂直分辨率绘制土壤水分图提供了参考。