Surface downward shortwave radiation (DSR) is a fundamental variable in determining the Earth's radiation balance and is essential in many applications. Considerable efforts have been devoted to algorithm development, product generation, and validation. However, few studies have focused on comparing retrieval approaches, examining their strengths and weaknesses, and identifying the most suitable scenarios for each approach. In this study, we implemented and evaluated five representative DSR retrieval algorithms, including the forward parameterization approach, two physical inversion methods (look-up table (LUT) and optimization), and two statistical inversion methods (direct estimation and neural networks). We then proposed an algorithm-integration framework that combined the results of these DSR retrieval methods to further improve DSR estimation accuracy and consistency. To validate the DSR retrievals, we used in-situ data collected at 25 stations of the Baseline Surface Radiation Network (BSRN) over one year. Validation revealed that forward parameterization consistently performed best, with an overall root mean square error (RMSE) of 91.7 W/m² or a relative RMSE of 16.9%, although it generated the fewest valid retrievals. For an identical data set, the LUT approach generated results comparable to those of parameterization. The neural network-based algorithm-integration approach reduced the RMSE by 11.0 W/m² or the relative RMSE by 2.0%, compared to the best individual retrieval algorithm. Our analysis demonstrates that algorithm integration is a promising way to obtain DSR data that are superior to estimates from any individual retrieval algorithm.

地表向下短波辐射 (DSR) 是确定地球辐射平衡的基本变量，在许多应用中都是必不可少的。大量的努力致力于算法开发、产品生成和验证。然而，很少有研究专注于比较检索方法，检查它们的优点和缺点，并为每种方法确定最合适的场景。在本研究中，我们实现并评估了五种具有代表性的 DSR 检索算法，包括前向参数化方法、两种物理反演方法（查找表 (LUT) 和优化）和两种统计反演方法（直接估计和神经网络）。然后，我们提出了一个算法集成框架，将这些 DSR 检索方法的结果结合起来，以进一步提高 DSR 估计的准确性和一致性。为了验证 DSR 反演，我们使用了一年多来在基线表面辐射网络 (BSRN) 的 25 个站点收集的原位数据。验证表明前向参数化始终表现最佳，总均方根误差 (RMSE) 为 91.7 W/m²或 16.9% 的相对 RMSE，尽管它生成的有效检索最少。对于相同的数据集，LUT 方法生成的结果与参数化的结果相当。与最佳个体检索算法相比，基于神经网络的算法集成方法将 RMSE 降低了 11.0 W/m ²或相对 RMSE 降低了 2.0%。我们的分析表明，算法集成是一种获得优于任何单个检索算法估计的 DSR 数据的有前途的方法。