A deep learning reconstruction of mass balance series for all glaciers in the French Alps: 1967-2015

Glacier surface mass balance (SMB) data are crucial to understand and quantify the regional effects of climate on glaciers and the high-mountain water cycle, yet observations cover only a small fraction of glaciers in the world. We present a dataset of annual glacier-wide surface mass balance of all the glaciers in the French Alps for the 1967-2015 period. This dataset has been reconstructed using deep learning (i.e. a deep artificial neural network), based on direct and remote sensing SMB observations, meteorological reanalyses and topographical data from glacier inventories. This data science reconstruction approach is embedded as a SMB component of the open-source ALpine Parameterized Glacier Model (ALPGM: https://zenodo.org/record/3609136). An extensive cross-validation allowed to assess the method’s validity, with an estimated average error (RMSE) of 0.49 m.w.e. a-1, an explained variance (r2) of 79% and an average bias of +0.017 m.w.e. a-1. We estimate an average regional area-weighted glacier-wide SMB of -0.72±0.20 m.w.e. a-1 for the 1967-2015 period, with moderately negative mass balances in the 1970s (-0.52 m.w.e. a-1) and 1980s (-0.12 m.w.e. a-1), and an increasing negative trend from the 1990s onwards, up to -1.39 m.w.e. a-1 in the 2010s. Following a topographical and regional analysis, we estimate that the massifs with the highest mass losses for this period are the Chablais (-0.90 m.w.e. a-1) and Ubaye and Champsaur (-0.91 m.w.e. a-1 both) ranges, and the ones presenting the lowest mass losses are the Mont-Blanc and Oisans ranges (-0.74 and -0.78 m.w.e. a-1 respectively). This dataset provides relevant and timely data for studies in the fields of glaciology, hydrology and ecology in the French Alps, in need of regional or glacier-specific meltwater contributions in glacierized catchments. The SMB dataset is comprised of multiple CSV files, one for each of the 661 glaciers from the 2003 glacier inventory (Gardent et al., 2014), named with its GLIMS ID and RGI ID with the following format: GLIMS-ID_RGI-ID_SMB.csv. Both indexes are used since some glaciers that split into multiple sub-glaciers do not have an RGI ID. Split glaciers have the GLIMS ID of their "parent" glacier and an RGI ID equal to 0. Every file contains one column for the year number between 1967 and 2015 and another column for the annual glacier-wide SMB time series. Glaciers with remote sensing-derived observations (Rabatel et al., 2016) include this information as an additional column. This allows the user to choose the source of data, with remote sensing data having lower uncertainties (0.35±0.06 () m.w.e. a-1 as estimated in Rabatel et al. (2016)). Columns are separated by semicolon (;). All topographical data for the 661 glaciers can be found in the updated version of the 2003 glacier inventory included in the Supplementary material.

冰川表面物质平衡（surface mass balance, SMB）数据对于理解并量化气候对冰川及高山区水循环的区域影响至关重要，但当前实地观测仅覆盖全球极小部分冰川。本研究构建了1967-2015年法国阿尔卑斯山区所有冰川的年度全域表面物质平衡数据集。

该数据集基于直接观测与遥感获取的SMB数据、气象再分析资料以及冰川编目中的地形数据，通过深度学习（即深度人工神经网络）方法重建得到。此数据科学重建方法被集成至开源阿尔卑斯参数化冰川模型（ALpine Parameterized Glacier Model, ALPGM，https://zenodo.org/record/3609136）的SMB组件中。

研究通过广泛的交叉验证评估了该方法的有效性，估计其平均误差（均方根误差, RMSE）为0.49 米水当量·年⁻¹（meter water equivalent per annum, m.w.e. a⁻¹），解释方差（r²）达79%，平均偏差为+0.017 米水当量·年⁻¹。

研究估算1967-2015年法国阿尔卑斯山区的区域面积加权全域SMB平均值为-0.72±0.20 米水当量·年⁻¹，其中20世纪70年代物质平衡为中度负值（-0.52 米水当量·年⁻¹），80年代为-0.12 米水当量·年⁻¹，90年代起负向趋势加剧，2010年代达到-1.39 米水当量·年⁻¹。

通过地形与区域分析，研究估算此时期内物质损耗最严重的山地块为夏布莱（Chablais）、于拜（Ubaye）和尚波（Champsaur）（三者均为-0.91 米水当量·年⁻¹），物质损耗最轻的为勃朗峰（Mont-Blanc）与瓦桑（Oisans）山地块（分别为-0.74与-0.78 米水当量·年⁻¹）。

该数据集可为法国阿尔卑斯山区的冰川学、水文学与生态学研究提供及时且相关的数据支持，相关研究往往需要冰川化流域的区域或单冰川融水贡献数据。

本SMB数据集包含多个逗号分隔值（Comma-Separated Values, CSV）文件，每个文件对应2003年冰川编目（Gardent等，2014）中的661条冰川之一，文件命名格式为：GLIMS-ID_RGI-ID_SMB.csv。其中GLIMS为全球冰川遥感测量数据库标识符（Global Land Ice Measurements from Space, GLIMS），RGI为伦道夫冰川编目标识符（Randolph Glacier Inventory, RGI）。同时使用两类标识符是因为部分分裂为多条子冰川的冰川无对应RGI ID。分裂冰川使用其“母冰川”的GLIMS ID，且RGI ID设为0。

每个文件包含两列基础数据：一列对应1967-2015年的年份，另一列对应年度全域SMB时间序列。带有遥感观测数据的冰川（Rabatel等，2016）会额外增加一列观测数据，方便使用者选择数据来源——遥感数据的不确定性更低，据Rabatel等（2016）估算为0.35±0.06 米水当量·年⁻¹。数据列以分号（;）分隔。

661条冰川的所有地形数据可在补充材料中收录的2003年冰川编目更新版本中获取。