Prior land surface reflectance-based sandstorm detection from space using deep learning

Traditional sandstorm detection methods use radiation differences among dust, underlying surface, and cloud to distinguish them by setting appropriate thresholds. Owing to the complex structure of the underlying surface, dust, and cloud, it is difficult to set a uniform threshold to achieve high-precision separation. Deep learning (DL) has powerful information mining capabilities and can fully use spectral differences between dust, land surface, and clouds. However, under the limited band information provided by satellite sensors, DL cannot easily distinguish highly heterogeneous land surfaces from multi-modal dust and cloud. This study proposes a sandstorm detection algorithm with DL supported by a land surface reflectance (LSR) dataset. The clear sky LSR dataset was obtained based on the MOD09A1 product. Based on the dataset, the difference between the reflectance observed by the satellite and the corresponding LSR is generated, which is used as a characteristic parameter of sandstorm detection with the deep learning method. The sandstorm detection of MODIS data is realized using multi-band radiation and radiation difference with DL. Results showed that the sandstorm detection algorithm used in this study was consistent with the OMI AI product with a detection accuracy of 84.6%. Compared with the detection results without the LSR dataset, this method effectively improves the accuracy of sandstorm identification.