黑河流域区域尺度地表蒸散发相对真值数据集（2012-2016年）ETMap Version 1.0

地表蒸散发（Evapotranspiration，ET）是连接着陆地能量平衡、水循环以及碳循环等的重要变量，地表蒸散发的准确获取有助于全球气候变化、作物估产、干旱监测等研究，并且对区域与全球的水资源规划管理具有重要的意义。地表蒸散发的获取方法主要包括地面观测、遥感估算、模式模拟与同化等。地面观测可以获得高精度的地表蒸散发数据，但观测站点的空间代表性十分有限；遥感估算、模式模拟与同化方法可以获得空间连续的地表蒸散发，但存在精度与时空分布格局合理性的验证问题。因此，本研究充分利用众多的高精度站点观测数据，结合多源遥感信息，将地面站点观测尺度扩展至区域上，获得高精度、时空分布连续的地表蒸散发量。
基于近年来开展的“黑河综合遥感联合试验”（WATER）、“黑河流域生态-水文过程综合遥感观测联合试验”（HiWATER）、所积累的站点观测数据（自动气象站、涡动相关仪、大孔径闪烁仪等），共选用36个站点（65个站年，分布图见图1），结合多源遥感数据（土地覆盖与植被类型图，叶面积指数、地表温度等）和大气驱动数据等，运用五种机器学习方法（回归树、随机森林、人工神经网络、支持向量机、深度信念网络）分别构建了不同的地表蒸散发尺度扩展模型，对各尺度扩展模型进行了全面的对比分析，结果表明：相比于其他四种方法，随机森林方法更适合于黑河流域由站点到区域的地表蒸散发尺度扩展研究。基于优选出的随机森林尺度扩展模型，以遥感及大气驱动数据作为输入，生产了2012~2016年生长季（5~9月）黑河流域地表蒸散发时空分布图（ETMap，时间分辨率为逐日，空间分辨率为1km）。以LAS观测值为真值进行验证，结果表明：ETMap整体精度良好，上游 (LAS1)、中游 (LAS2-LAS5)和下游 (LAS6 - LAS8)的RMSE (MAPE)分别为0.65 mm/day(18.86%)、0.99 mm/day (19.13%)和0.91 mm/day (22.82%)。总之，ETMap是基于站点观测数据运用随机森林算法进行尺度扩展得到的精度较高的黑河流域地表蒸散发产品。所有站点信息和尺度扩展方法请参考Xu et al. (2018)，观测数据处理请参考Liu et al. (2018)。

Evapotranspiration (ET) is a vital variable that connects terrestrial energy balance, water cycle, and carbon cycle, among other important Earth system processes. Accurate retrieval of ET is essential for research on global climate change, crop yield estimation, drought monitoring, and other related fields, and holds great significance for regional and global water resource planning and management. The main approaches for obtaining ET include ground-based observation, remote sensing estimation, model simulation and data assimilation, among others. Ground-based observation can acquire high-precision ET data, but the spatial representativeness of observation stations is extremely limited; remote sensing estimation, model simulation and data assimilation methods can produce spatially continuous ET products, but they face challenges in verifying the accuracy and rationality of their spatiotemporal distribution patterns. Therefore, this study makes full use of a large number of high-precision station observation data, combined with multi-source remote sensing information, to scale up ground-based station observations to the regional scale, thereby obtaining high-precision, spatially and temporally continuous ET products.

Based on the accumulated station observation data (including automatic weather stations, eddy covariance systems, large-aperture scintillometers, etc.) from the recently conducted Heihe Integrated Remote Sensing Joint Experiment (WATER) and the Heihe River Basin Eco-Hydrological Process Integrated Remote Sensing Observation Joint Experiment (HiWATER), a total of 36 stations (65 station-years, with their distribution shown in Figure 1) were selected. Combined with multi-source remote sensing data (such as land cover and vegetation type maps, leaf area index, surface temperature, etc.) and atmospheric forcing data, five machine learning methods (regression tree, random forest, artificial neural network, support vector machine, deep belief network) were used to separately develop different ET scaling models. A comprehensive comparative analysis was conducted for each scaling model, and the results showed that compared with the other four methods, the random forest method is more suitable for the ET scaling research from station to regional scale in the Heihe River Basin. Based on the optimized random forest scaling model, using remote sensing and atmospheric forcing data as inputs, we produced the spatiotemporal distribution map of ET in the Heihe River Basin during the growing seasons (May to September) from 2012 to 2016 (ETMap, with a temporal resolution of daily and a spatial resolution of 1 km). Validation using LAS observations as the ground truth showed that ETMap has good overall accuracy. The RMSE (MAPE) values for the upstream (LAS1), midstream (LAS2-LAS5), and downstream (LAS6-LAS8) regions are 0.65 mm/day (18.86%), 0.99 mm/day (19.13%), and 0.91 mm/day (22.82%), respectively. In summary, ETMap is a high-precision ET product for the Heihe River Basin, which was generated by scaling up station observation data using the random forest algorithm. Please refer to Xu et al. (2018) for all station information and scaling methods, and Liu et al. (2018) for the observation data processing.