Abstract For the past two decades, quantitative retrievals of aerosol optical depth (AOD) have been made from both geostationary and polar-orbiting satellites, and the results have been widely used in numerous studies. Despite the progress made in improving the accuracy of AOD retrievals, there are still major challenges, especially over land. A notable one for the so-called Dark-Target (DT) algorithms is building the surface reflectance (SR) relationships (SRR) to derive SR in the visible channels from SR in the short-wave infrared (SWIR) channel, mainly because these relationships are strongly subjected to entangled factors (e.g., viewing geometry, surface type, and vegetation state). In this study, we examine the benefits of a new method for deriving the SRR using deep learning techniques. The SRR constructed by the deep neural network (DNN) considers multiple related inputs, such as the SWIR normalized difference vegetation index (NDVISWIR), viewing geometry, and seasonality, among others. We then incorporate the DNN-constrained SRR into a DT algorithm developed at NOAA/STAR to retrieve AOD from the Advanced Himawari Instrument (AHI) onboard the new generation of geostationary satellites, Himawari-8. The revised DT algorithm with the deep learning technique (DTDL) demonstrates improved performance over the study region (95–125°E, 18–30°N, a portion of the AHI full disk), as attested by significantly reduced random noise, especially for low NDVISWIR and high surface albedo cases. Robust independent tests indicate that this algorithm can be applied to untrained regions, not only to those used in training. The method directly benefits the algorithm development for Himawari-8 and can also be adopted for other geostationary or polar-orbiting satellites. Our study illustrates how artificial intelligence could significantly improve AOD retrievals from multi-spectral satellite observations following this new approach.

中文翻译：

---

通过深度学习构建光谱表面反射率的关系，完善陆地上的气溶胶光学深度反演：在 Himawari-8 中的应用

摘要 近二十年来，地球同步卫星和极轨卫星都对气溶胶光学深度（AOD）进行了定量反演，其结果被广泛应用于众多研究中。尽管在提高 AOD 检索准确性方面取得了进展，但仍然存在重大挑战，尤其是在陆地上。所谓的暗目标 (DT) 算法的一个值得注意的是建立表面反射 (SR) 关系 (SRR)，以从短波红外 (SWIR) 通道中的 SR 中导出可见光通道中的 SR，主要是因为这些关系强烈地受到纠缠因素的影响（例如，观察几何形状、表面类型和植被状态）。在这项研究中，我们研究了一种使用深度学习技术推导 SRR 的新方法的好处。由深度神经网络 (DNN) 构建的 SRR 考虑了多个相关输入，例如 SWIR 归一化差异植被指数 (NDVISWIR)、查看几何和季节性等。然后，我们将受 DNN 约束的 SRR 合并到 NOAA/STAR 开发的 DT 算法中，以从新一代地球同步卫星 Himawari-8 上的高级 Himawari Instrument (AHI) 中检索 AOD。使用深度学习技术 (DTDL) 修改后的 DT 算法在研究区域（东经 95-125°，北纬 18-30°，AHI 完整磁盘的一部分）上表现出改进的性能，这可以通过显着降低随机噪声来证明，尤其是适用于低 NDVISWIR 和高表面反照率情况。稳健的独立测试表明，该算法可以应用于未训练的区域，而不仅仅是用于训练的区域。该方法直接有利于 Himawari-8 的算法开发，也可用于其他地球静止或极轨卫星。我们的研究说明了人工智能如何通过这种新方法显着改善多光谱卫星观测的 AOD 检索。