Abstract. The surface radiation budget, also known as all-wave net radiation (R_n), is a key parameter for various land surface processes including hydrological, ecological, agricultural, and biogeochemical processes. Satellite data can be effectively used to estimate R_n, but existing satellite products have coarse spatial resolutions and limited temporal coverage. In this study, a point-surface matching estimation (PSME) method is proposed to estimate surface R_n using a residual convolutional neural network (RCNN) integrating spatially adjacent information to improve the accuracy of retrievals. A global high-resolution (0.05°) long-term (1981–2019) R_n product was subsequently generated from Advanced Very High-Resolution Radiometer (AVHRR) data. Specifically, the RCNN was employed to establish a nonlinear relationship between globally distributed ground measurements from 537 sites and AVHRR top of atmosphere (TOA) observations. Extended triplet collocation (ETC) technology was applied to address the spatial scale mismatch issue resulting from the low spatial support of ground measurements within the AVHRR footprint by selecting reliable sites for model training. The overall independent validation results show that the generated AVHRR R_n product is highly accurate, with R², root-mean-square error (RMSE), and bias of 0.84, 26.66 Wm⁻² (31.66 %), and 1.59 Wm⁻² (1.89 %), respectively. Inter-comparisons with three other R_n products, i.e., the 5 km Global Land Surface Satellite (GLASS), the 1° Clouds and the Earth's Radiant Energy System (CERES), and the 0.5° × 0.625° Modern-Era Retrospective analysis for Research and Applications, Version 2 (MERRA2), illustrate that our AVHRR R_n retrievals have the best accuracy under all of the considered surface and atmospheric conditions, especially thick cloud or hazy conditions. The spatiotemporal analyses of these four R_n datasets indicate that the AVHRR R_n product reasonably replicates the spatial pattern and temporal evolution trends of R_n observations. This dataset is freely available at https://doi.org/10.5281/zenodo.5509854 for 1981–2019 (Xu et al., 2021).

中文翻译：

---

使用残差卷积神经网络从 AVHRR 卫星数据获得的全球长期（1981-2019）每日地表辐射收支产品

摘要。地表辐射收支，也称为全波净辐射 (R _n )，是包括水文、生态、农业和生物地球化学过程在内的各种地表过程的关键参数。卫星数据可以有效地用于估计 R _n，但现有的卫星产品空间分辨率粗糙，时间覆盖范围有限。在这项研究中，提出了一种点表面匹配估计（PSME）方法来估计表面 R _n使用残差卷积神经网络（RCNN）集成空间相邻信息以提高检索的准确性。全球高分辨率 (0.05°) 长期 (1981–2019) R _n产品随后由高级甚高分辨率辐射计 (AVHRR) 数据生成。具体而言，RCNN 用于建立来自 537 个站点的全球分布式地面测量值与 AVHRR 大气顶（TOA）观测值之间的非线性关系。通过选择可靠的模型训练站点，扩展三重配置 (ETC) 技术被应用于解决由于 AVHRR 足迹内地面测量的空间支持低而导致的空间尺度不匹配问题。整体独立验证结果表明，生成的 AVHRR R _n产品高度准确，R ²、均方根误差 (RMSE) 和偏差为 0.84、26.66 Wm ^-2 (31.66 %) 和 1.59 Wm ^-2(1.89%)，分别。与其他三个 R _n产品的相互比较，即 5 公里全球陆地表面卫星 (GLASS)、1°云和地球辐射能系统 (CERES)，以及 0.5° × 0.625° 现代时代回顾性分析研究与应用，第 2 版 (MERRA2) 表明，我们的 AVHRR R _n反演在所有考虑的地表和大气条件下，尤其是厚云或朦胧条件下具有最佳精度。这四个R的时空分析_Ñ数据集指示该AVHRRř _Ñ产物合理复制R的空间图案和时间演进趋势_Ñ观察。该数据集可在 https://doi.org/10.5281/zenodo.5509854 上免费获得 1981-2019 年（Xu 等人，2021）。