Research on infrared hyperspectral remote sensing cloud detection method based on deep learning

Infrared hyperspectral is susceptible to clouds, and accurately identifying whether the hyperspectral infrared sounder data is polluted by clouds is of great significance for numerical weather prediction and atmospheric parameter inversion. Since the complex spectral characteristics of clouds, the existing spectral threshold methods and machine learning methods have the difficulties of undetermined threshold and clear field of view (FOV) missed and false detections. In order to improve the cloud recognition accuracy of infrared hyperspectral data, three end-to-end cloud detection models combining deep neural network (DNN) and convolutional neural network (CNN) and long short-term memory network (LSTM) are proposed. In this paper, taking the High Spectral Infrared Atmospheric Sounder (HIRAS) equipped with Fengyun-3D (FY-3D) satellite as the research object, based on the same platform Moderate Resolution Spectral Imager-II (MERSI-II) cloud mask (CLM) product, the HIRAS Cloud dataset is established, and the accuracy test and qualitative analysis are carried out by using the test datasets and Typhoon Siamba, July 3, 2022, as well as the earth observation scene under the conditions of ice and snow surface. The test datasets analysis results show that the cloud detection accuracy of CNN and CNN-LSTM model is stable at 0.96, and the false alarm rate of cloud is 0.035 and 0.036, respectively, and the detection ability of DNN model is slightly inferior to the former two in the same hidden layer, with an accuracyof 0.94. In further qualitative research, we found that the CNN-LSTM model has high accuracy and robustness in infrared hyperspectral cloud detection, and the detection results in a variety of surface scenarios are consistent with the actual situation of whether clouds occur in the FOV of the instrument. Compared with CLM products, it can better identify clear ocean scenes, and provide fast and efficient cloud detection reference for data assimilation systems.

Automated summary

Infrared hyperspectral data is easily affected by clouds, and accurately identifying cloud pollution is important for weather prediction and parameter inversion. To improve cloud recognition, three end-to-end cloud detection models combining DNN, CNN, and LSTM are proposed. The models are tested using datasets and various surface scenarios, showing high accuracy and consistency with actual cloud occurrences. Compared to existing products, the models can better identify clear ocean scenes and provide efficient cloud detection reference for data assimilation systems.