Snow water equivalent (SWE) is among the most important variables in the hydrological modelling of high latitude and mountainous areas. While manual snow surveys can directly provide SWE measurements, they are time consuming and costly, especially compared to automated snow depth measurements. Moreover, SWE is strongly correlated to snow depth. For this reason, several empirical equations relating snow depth to SWE have been proposed. The present study investigates the potential of artificial neural networks for estimating SWE from snow depth and commonly available data, and the proposed method is compared to existing, regression-based methods. An ensemble of multilayer perceptrons is constructed and trained using gridded meteorological variables and a data set of almost 40,000 SWE and depth measurements from the province of Quebec (eastern Canada). Overall, the proposed artificial neural network-based method reached a RMSE of 28 mm and outperforms by 17% a series of empirical equations for estimating the SWE of an independent set of measurement sites. Nevertheless, all the tested methods demonstrated limits to estimate lowest values of snow bulk density.

在高纬度和山区的水文模拟中，雪水当量（SWE）是最重要的变量之一。尽管手动降雪测量可以直接提供SWE测量，但它们既费时又昂贵，特别是与自动降雪深度测量相比。此外，SWE与雪深密切相关。因此，提出了一些将雪深与SWE相关的经验公式。本研究调查了人工神经网络从降雪深度和常用数据估算SWE的潜力，并将该方法与现有的基于回归的方法进行了比较。使用网格化的气象变量和将近40个数据集来构造和训练多层感知器 来自魁北克省（加拿大东部）的000 SWE和深度测量。总体而言，所提出的基于人工神经网络的方法达到了28 mm的均方根误差（RMSE），并且比一系列经验方程式的效果要高出17％，这些经验方程式用于估算一组独立的测量站点的SWE。然而，所有测试方法都显示出估计最低积雪密度值的限制。