准确评估雪层体积含水量和体积密度对于理解雪的水文过程、降低雪崩风险以及冰冻圈监测具有重要意义。本文提出一种新型双参数反演框架,该框架集成了合成电磁建模、降维方法和机器学习算法,用于从探地雷达(GPR)数据中提取相对介电常数...准确评估雪层体积含水量和体积密度对于理解雪的水文过程、降低雪崩风险以及冰冻圈监测具有重要意义。本文提出一种新型双参数反演框架,该框架集成了合成电磁建模、降维方法和机器学习算法,用于从探地雷达(GPR)数据中提取相对介电常数和对数电阻率。传统雪层测量方法具有侵入性、劳动强度大且仅限于点位观测等局限性。为克服上述局限,建立了一种非侵入性、可扩展且数据驱动的框架,利用合成GPR数据集来表示具有不同含水量和密度分布的多样化雪层条件。使用先进电磁模拟器gprMax,通过有限差分时域模拟生成合成的一维时序反射(A扫描)。随后采用主成分分析(PCA)将每个A扫描进行压缩,得到低维且信息保真的特征集,从而显著提升模型训练效率。基于经过主成分分析(PCA)降维处理的特征集,训练了随机森林、神经网络、支持向量机和极限梯度提升四种机器学习模型。其中,神经网络模型性能最佳,介电常数为R^(2)>0.97,电阻率为R^(2)>0.92。合成数据中引入高斯噪声(信噪比约为6 dB),并通过针对特定领域进行改进,以提高其在实地数据的泛化能力。模型在中国阿尔泰山脉的两条典型GPR剖面(湿雪T750和饱和雪G125)上进行了验证。神经网络模型预测结果与GPR反演、Snowfork测量及人工雪坑数据高度一致,体积含水量偏差不超过1.5%,体积密度误差在30–84 kg m-3范围内。结果表明,基于机器学习的反演方法在结合真实模拟与数据增强的条件下,能够实现可扩展、非侵入式的雪层特性反演,在水文预报、雪情监测及气候敏感型水资源管理中具有重要应用价值。展开更多
Accurate assessment of snowpack volumetric liquid water content and bulk density is essential for understanding snow hydrology,avalanche risk management,and monitoring cryosphere changes.This study presents a novel du...Accurate assessment of snowpack volumetric liquid water content and bulk density is essential for understanding snow hydrology,avalanche risk management,and monitoring cryosphere changes.This study presents a novel dual-parameter inversion framework that integrates synthetic electromagnetic modelling,dimensionality reduction,and machine learning algorithms to extract relative permittivity and log-resistivity from ground-penetrating radar(GPR)data.Traditional snowpack measurements are invasive,labor-intensive,and limited to point observations.To overcome these limitations,we developed a non-invasive,scalable,and data-driven framework that uses synthetic GPR datasets representing diverse snowpack conditions with variable moisture and density profiles.Synthetic 1D time series reflections(A-scans)are generated using finite-difference time-domain simulations in the state-of-the-art electromagnetic simulator gprMax.Principal component analysis(PCA)is applied to compress each A-scan while preserving key features,which significantly improved and enhanced the model training efficiency.Four machine learning models,including random forest,neural network,support vector machine,and eXtreme gradient boosting,are trained on PCA-reduced features.Among these,the neural network model achieved the best performance,with R^(2)>0.97 for permittivity and R 2>0.92 for resistivity.Gaussian noise(signal-to-noise ratio of 6 dB)is introduced to the synthetic data,and then targeted domain adaptation is employed to enhance generalization to field data.The framework is validated on two contrasting GPR transects in the Altay Mountains of the Chinese mainland,representing moist(T750)and wet(G125)snowpack conditions.The neural network model predictions are most consistent with the GPR derived estimates,Snowfork measurements,and snow pit data,achieving volumetric liquid water content deviation of≤1.5% and bulk density error within the range of 30-84 kg m^(-3).The results demonstrate that machine learning-based inversion,supported by realistic simulations and data augmentation enables scalable,non-invasive snowpack characterization with significant applications in hydrological forecasting,snow monitoring,and water resource management.展开更多
文摘准确评估雪层体积含水量和体积密度对于理解雪的水文过程、降低雪崩风险以及冰冻圈监测具有重要意义。本文提出一种新型双参数反演框架,该框架集成了合成电磁建模、降维方法和机器学习算法,用于从探地雷达(GPR)数据中提取相对介电常数和对数电阻率。传统雪层测量方法具有侵入性、劳动强度大且仅限于点位观测等局限性。为克服上述局限,建立了一种非侵入性、可扩展且数据驱动的框架,利用合成GPR数据集来表示具有不同含水量和密度分布的多样化雪层条件。使用先进电磁模拟器gprMax,通过有限差分时域模拟生成合成的一维时序反射(A扫描)。随后采用主成分分析(PCA)将每个A扫描进行压缩,得到低维且信息保真的特征集,从而显著提升模型训练效率。基于经过主成分分析(PCA)降维处理的特征集,训练了随机森林、神经网络、支持向量机和极限梯度提升四种机器学习模型。其中,神经网络模型性能最佳,介电常数为R^(2)>0.97,电阻率为R^(2)>0.92。合成数据中引入高斯噪声(信噪比约为6 dB),并通过针对特定领域进行改进,以提高其在实地数据的泛化能力。模型在中国阿尔泰山脉的两条典型GPR剖面(湿雪T750和饱和雪G125)上进行了验证。神经网络模型预测结果与GPR反演、Snowfork测量及人工雪坑数据高度一致,体积含水量偏差不超过1.5%,体积密度误差在30–84 kg m-3范围内。结果表明,基于机器学习的反演方法在结合真实模拟与数据增强的条件下,能够实现可扩展、非侵入式的雪层特性反演,在水文预报、雪情监测及气候敏感型水资源管理中具有重要应用价值。
基金supported by the National Key R&D Program of China(Grant Nos.2023YFC3008300&2023YFC3008305)the National Natural Science Foundation of China(Grant No.42172320)+1 种基金the Key Laboratory of Mountain Hazards and Engineering Resilience,Institute of Mountain Hazards and Environment,Chinese Academy of Sciences(Grant Nos.KLMHER-Z06&KLMHER-T07)the Science and Technology Research Program of Institute of Mountain Hazards and Environment,Chinese Academy of Sciences(Grant No.IMHE-CXTD.04).
文摘Accurate assessment of snowpack volumetric liquid water content and bulk density is essential for understanding snow hydrology,avalanche risk management,and monitoring cryosphere changes.This study presents a novel dual-parameter inversion framework that integrates synthetic electromagnetic modelling,dimensionality reduction,and machine learning algorithms to extract relative permittivity and log-resistivity from ground-penetrating radar(GPR)data.Traditional snowpack measurements are invasive,labor-intensive,and limited to point observations.To overcome these limitations,we developed a non-invasive,scalable,and data-driven framework that uses synthetic GPR datasets representing diverse snowpack conditions with variable moisture and density profiles.Synthetic 1D time series reflections(A-scans)are generated using finite-difference time-domain simulations in the state-of-the-art electromagnetic simulator gprMax.Principal component analysis(PCA)is applied to compress each A-scan while preserving key features,which significantly improved and enhanced the model training efficiency.Four machine learning models,including random forest,neural network,support vector machine,and eXtreme gradient boosting,are trained on PCA-reduced features.Among these,the neural network model achieved the best performance,with R^(2)>0.97 for permittivity and R 2>0.92 for resistivity.Gaussian noise(signal-to-noise ratio of 6 dB)is introduced to the synthetic data,and then targeted domain adaptation is employed to enhance generalization to field data.The framework is validated on two contrasting GPR transects in the Altay Mountains of the Chinese mainland,representing moist(T750)and wet(G125)snowpack conditions.The neural network model predictions are most consistent with the GPR derived estimates,Snowfork measurements,and snow pit data,achieving volumetric liquid water content deviation of≤1.5% and bulk density error within the range of 30-84 kg m^(-3).The results demonstrate that machine learning-based inversion,supported by realistic simulations and data augmentation enables scalable,non-invasive snowpack characterization with significant applications in hydrological forecasting,snow monitoring,and water resource management.
