Glaciers play a vital role in providing water resources for drinking, agriculture, and hydro-electricity in many mountainous regions. As global warming progresses, accurately reconstructing long-term glacier mass changes and comprehending their intricate dynamic relationships with environmental variables are imperative for sustaining livelihoods in these regions. This paper presents the use of eXplainable Machine Learning (XML) models with GRACE and GRACE-FO data to reconstruct long-term monthly glacier mass changes in the Upper Yukon Watershed (UYW), Canada. We utilized the H2O-AutoML regression tools to identify the best performing Machine Learning (ML) model for filling missing data and predicting glacier mass changes from hydroclimatic data. The most accurate predictive model in this study, the Gradient Boosting Machine, coupled with explanatory methods based on SHapley Additive eXplanation (SHAP) and Local Interpretable Model-Agnostic Explanations (LIME) analyses, led to automated XML models. The XML unveiled and ranked key predictors of glacier mass changes in the UYW, indicating a decrease since 2014. Analysis showed decreases in snow water equivalent, soil moisture storage, and albedo, along with increases in rainfall flux and air temperature were the main drivers of glacier mass loss. A probabilistic analysis hinging on these drivers suggested that the influence of the key hydrological features is more critical than the key meteorological features. Examination of climatic oscillations showed that high positive anomalies in sea surface temperature are correlated with rapid depletion in glacier mass and soil moisture, as identified by XML. Integrating H2O-AutoML with SHAP and LIME not only achieved high prediction accuracy but also enhanced the explainability of the underlying hydroclimatic processes of glacier mass change reconstruction from GRACE and GRACE-FO data in the UYW. This automated XML framework is applicable globally, contingent upon sufficient high-quality data for model training and validation.

中文翻译：

---

使用可解释的自动化机器学习算法解释 GRACE 对育空上游流域内冰川质量变化的解释


冰川在许多山区为饮用水、农业和水电提供水资源方面发挥着至关重要的作用。随着全球变暖的进展，准确重建冰川的长期质量变化并理解它们与环境变量的复杂动态关系对于维持这些地区的生计至关重要。本文介绍了使用 eXplainable 机器学习 （XML） 模型与 GRACE 和 GRACE-FO 数据来重建加拿大育空河流域 （UYW） 的长期月度冰川质量变化。我们利用 H2O-AutoML 回归工具来确定性能最佳的机器学习 （ML） 模型，用于填充缺失数据并根据水文气候数据预测冰川质量变化。本研究中最准确的预测模型，梯度提升机，结合基于 SHapley 加性解释 （SHAP） 和局部可解释模型不可知解释解释 （LIME） 分析的解释方法，导致了自动化 XML 模型。XML 揭示了 UYW 中冰川质量变化的关键预测因子并对其进行了排名，表明自 2014 年以来有所下降。分析表明，雪水当量、土壤水分储存和反照率的减少，以及降雨通量和空气温度的增加是冰川质量损失的主要驱动因素。基于这些驱动因素的概率分析表明，关键水文特征的影响比关键气象特征的影响更关键。对气候振荡的检查表明，正如 XML 所确定的那样，海面温度的高正异常与冰川质量和土壤水分的快速消耗相关。 将 H2O-AutoML 与 SHAP 和 LIME 集成不仅实现了高预测精度，而且还增强了 UYW 中 GRACE 和 GRACE-FO 数据对冰川质量变化重建的潜在水文气候过程的可解释性。这种自动化 XML 框架在全球范围内适用，具体取决于用于模型训练和验证的足够高质量数据。