Abstract Snow cover models such as Crocus or S nowpack have been designed to simulate the detailed stratigraphy of snow properties. This is relevant, for instance, to assess snowpack stability in support of avalanche forecasting. However, such models have generally been evaluated on bulk or surface properties, such as snow depth, water equivalent of snow cover or surface albedo, but not on the detailed stratigraphy. The large number of snow profiles collected in observer networks have thus not been assimilated in such models hitherto. This study introduces a new method to (1) directly compare simulated and observed snow layering and (2) allow for the insertion of observed profile to initialize a snow cover model. This method is mainly based on a scheme to convert observations into state variables of snow cover models and matching observed and simulated layering, accounting for potential depth shifts. The developed methodology was applied to the Crocus snow cover model at three sites in the French Alps, for 15 winter seasons between 2000 and 2015. The performance of Crocus initialized with a bare ground at the end of the summer was evaluated against 739 observed profiles. The model performance varied with the considered winter season and sites. On average, Crocus reproduced snow depth with a median error of 12 cm, layer density with a median error of 50 kg m−3, layer grain shape with an error of 0.31 according to a specially developed metric. The re-initialization of the model with observed profiles during winter season enabled to reduce these simulation errors. One week after the direct insertion of a manual profile, the median error of the simulation decreased to 6.8 cm for snow depth, 39 kg m−3 for density and 0.25 for grain shape. However, the improvement provided by this re-initialization almost completely vanished one month after the insertion.

中文翻译：

---

通过地层观测强制和评估详细的积雪模型

摘要 积雪模型（例如 Crocus 或 S nowpack）旨在模拟积雪特性的详细地层。例如，这与评估积雪稳定性以支持雪崩预测有关。然而，这些模型通常是根据整体或表面特性进行评估的，例如雪深、积雪的水当量或地表反照率，而不是详细地层。因此，迄今为止在此类模型中尚未同化在观测器网络中收集的大量雪剖面。本研究引入了一种新方法来 (1) 直接比较模拟和观察到的积雪分层和 (2) 允许插入观察到的剖面来初始化积雪模型。该方法主要基于将观测转换为积雪模型的状态变量并匹配观测和模拟分层的方案，考虑潜在的深度偏移。开发的方法应用于法国阿尔卑斯山三个地点的 Crocus 积雪模型，在 2000 年至 2015 年之间的 15 个冬季。根据 739 个观察到的配置文件评估了在夏末以裸露地面初始化的 Crocus 的性能。模型性能因考虑的冬季和地点而异。平均而言，根据专门开发的度量标准，Crocus 再现了雪深的中值误差为 12 cm，层密度的中值误差为 50 kg m-3，层粒形状的误差为 0.31。使用冬季观测到的剖面重新初始化模型能够减少这些模拟错误。直接插入手动配置文件一周后，雪深的模拟误差中位数降至 6.8 cm，密度为 39 kg m-3，颗粒形状为 0.25。然而，这种重新初始化提供的改进在插入后一个月几乎完全消失。