高精度沙尘检测对沙尘传输研究、气候评估及大气环境研究具有重要意义。本研究基于国产FY?4A卫星搭载的先进的静止轨道辐射成像仪（AGRI）观测数据，结合ERA?5再分析资料等数据，提出了一种融合空谱注意力机制的改进U?Net沙尘检测模型（SA?UNet）。在模型输入方面，考虑了沙尘的光谱特性与空间分布特征，构建了包含多波段反射率、亮温观测、沙尘光谱指数以及气象参数的多维tile级特征输入。该模型通过在U?Net架构中引入通道和空间双注意力模块，显著提升了模型对沙尘空间和光谱特征的提取能力；采用残差跳跃连接结构，有效缓解深层网络的梯度退化问题。验证结果表明，SA?UNet在测试集上的整体准确率达到99.46%，平均交并比为93.57%，F1-score为87.17%，验证精度显著优于传统的物理指数方法和仅依赖空间信息的机器学习方法。在独立沙尘事件检测测试中（数据未参与训练），SA?UNet的沙尘识别结果与目视解译结果的总体准确率达到98.77%，进一步验证了模型泛化能力。本研究可为基于FY?4A观测数据的大范围沙尘检测和传输监测提供技术参考。

Objective: Dust aerosols play a critical role in atmospheric radiative forcing, climate change, and air quality. Accurate dust detection is essential for dust transport studies and environmental monitoring. However, traditional dust detection methods—such as threshold-based physical indices or pure spatial machine learning models—are often limited in capturing the spectral spatial features of dust, especially over heterogeneous land surfaces. To address this challenge, this study proposes an improved U Net dust detection model that integrates a spatial spectral attention mechanism (SA UNet). The objective is to enhance the detection of dust events using FY-4A/AGRI observations, and to provide a reliable and generalizable method for large scale operational dust monitoring. Method: The proposed SA UNet is built upon a U Net architecture with two key innovations. First, a dual attention module combining channel and spatial attention is embedded into the encoder–decoder structure to adaptively emphasize informative spectral bands and crucial spatial regions related to dust plumes. Second, residual skip connections replace plain skip connections to mitigate gradient degradation in deep networks, improving training stability and feature reuse. For model inputs, we construct a multi dimensional “tile scale” feature set that includes multi band reflectance and brightness temperature from the Advanced Geosynchronous Radiation Imager (AGRI) onboard FY-4A, dust sensitive spectral indices (e.g., brightness temperature difference), and auxiliary meteorological parameters from ERA 5 reanalysis data (e.g., temperature, humidity, and wind fields). All features are normalized and organized into spatial tiles that capture both local details and the broader dust distribution context. The training dataset consists of manually labeled dust and non dust samples from diverse seasons and surface types. The model is trained using a loss function of binary cross entropy. Result: Quantitative evaluation on the independent test set shows that SA UNet achieves an overall accuracy of 99.46%, a mean intersection over union (mIoU) of 93.57%, and an F1 score of 87.17%. These metrics significantly outperform traditional physical index methods (e.g., brightness temperature difference approaches) and conventional machine learning models that rely only on spatial information. To further assess the generalization capability, a separate validation experiment was conducted on an independent dust event dataset that was not included in model training. In this challenge test, the dust identification results from SA UNet were compared with manual visual interpretation, yielding an overall accuracy of 98.77%. This high consistency confirms that the model does not simply memorize training samples but learns physically meaningful spatial spectral features of dust aerosols. Conclusion: In summary, this study demonstrates that integrating a spatial spectral attention mechanism into a deep learning framework substantially improves the accuracy and robustness of satellite based dust detection. The SA UNet effectively overcomes the limitations of threshold based indices and spatially only models by jointly exploiting the rich spectral information and spatial patterns of dust as observed by FY-4A/AGRI. The proposed method provides a reliable technical reference for operational, large scale dust monitoring and dust transport analysis over China and surrounding regions. It also shows promise for extension to other geostationary satellite sensors and aerosol type classification tasks.