Landsat imagery offers remarkable potential for various applications, including land monitoring and environmental assessment, thanks to its high spatial resolution and over 50 years of data records. However, the presence of atmospheric aerosols greatly hinders the precision of land classification and the quantitative retrieval of surface parameters. There is a pressing need for reliable and accurate global aerosol optical depth (AOD) data derived from Landsat imagery, particularly for atmospheric correction purposes and various other applications. To address this issue, we introduce an innovative framework for retrieving AOD from Landsat imagery over land, which leverages the deep-learning Transformer model (named AeroTrans-Landsat) and operates on the Google Earth Engine (GEE) cloud platform. We gather Landsat 8 and 9 images starting from their launch dates (February 2013 and September 2021, respectively) until the end of 2022, which are used to construct a robust aerosol retrieval model. The global AOD retrievals are then rigorously validated across ∼560 monitoring stations on land using diverse spatiotemporally independent methods. Leveraging information from multiple spectral channels, which contributes to 80 % according to the SHapley Additive exPlanation (SHAP) method, our retrieved AODs from 2013 to 2022 generally agree well with surface observations, with a sample-based cross-validation correlation coefficient of 0.905 and a root-mean-square error of 0.083. Around 86 % and 55 % of our AOD retrievals meet the criteria of Moderate Resolution Imaging Spectroradiometer (MODIS) Deep Blue expected errors [±(0.05 + 20 %)] and the Global Climate Observation System {[max(0.03, 10 %)]}, respectively. Additionally, our model is not as sensitive to fluctuations in both surface and atmospheric conditions, enabling the generation of spatially continuous AOD distributions with exceptionally fine-scale information over dark to bright surfaces. This capability extends to areas characterized by high pollution levels originating from both anthropogenic and natural sources.

中文翻译：

---

通过集成 Transformer 和 Google Earth Engine 的 Landsat 图像检索陆地上的全球气溶胶


由于其高空间分辨率和 50 多年的数据记录，陆地卫星图像为各种应用提供了巨大的潜力，包括土地监测和环境评估。然而，大气气溶胶的存在极大地阻碍了土地分类的精度和地表参数的定量反演。迫切需要从陆地卫星图像中获取可靠且准确的全球气溶胶光学深度 (AOD) 数据，特别是用于大气校正目的和各种其他应用。为了解决这个问题，我们引入了一种创新框架，用于从陆地上的 Landsat 图像中检索 AOD，该框架利用深度学习 Transformer 模型（名为 AeroTrans-Landsat）并在 Google Earth Engine (GEE) 云平台上运行。我们收集了 Landsat 8 和 9 从发射日期（分别为 2013 年 2 月和 2021 年 9 月）到 2022 年底的图像，这些图像用于构建强大的气溶胶反演模型。然后，使用不同的时空独立方法，在陆地上约 560 个监测站对全球 AOD 反演进行严格验证。利用来自多个光谱通道的信息（根据 SHapley 加法解释 (SHAP) 方法，这些信息占 80%），我们检索到的 2013 年至 2022 年 AOD 总体上与表面观测结果吻合良好，基于样本的交叉验证相关系数为 0.905，均方根误差为 0.083。我们约 86 % 和 55 % 的 AOD 反演符合中分辨率成像光谱仪 (MODIS) 深蓝预期误差 [±(0.05 + 20 %)] 和全球气候观测系统 {[max(0.03, 10 %)] 的标准}， 分别。 此外，我们的模型对地表和大气条件的波动不太敏感，能够生成空间连续的 AOD 分布，并在暗到亮的表面上提供极其精细的尺度信息。这种能力扩展到人为和自然来源污染水平较高的地区。