Cloud detection from the hyperspectral infrared radiation using random forest model

To use infrared observation data from the High Spectral Infrared Atmospheric Sounder (HIRAS) which is onboard the FengYun 3D (FY-3D) satellite, we have proposed a new cloud detection method based on random forest (RF). The true cloud distribution of field of views (FOVs) is generated by the collocated cloud masks of the Medium Resolution Spectral Imager-II (MERSI). The long-wave infrared radiations of 781 channels in the HIRAS FOVs are used as the input features of the model. The matched observation data of HIRAS and MERSI in East Asia (May 2019 to April 2020) are used as training and testing datasets. Given the significant differences in the radiation characteristics between land and sea, we respectively build the sea and land cloud detection models based on random forest. Both of them have achieved good cloud detection performance. The sea model produced slightly higher performance (ACC of 0.96, a FAR of 0.03, an F1-score of 0.96, and AUC of 0.99) than the land model (ACC of 0.95, FAR of 0.04, F1-score of 0.96, and AUC of 0.99). The RF cloud detection models have adequate generalization performances for the observations of HIRAS at different times and regions. Besides, the RF cloud detection models have faster computing efficiency and lower data dependency than HIRAS-MERSI matching method. The validation experiments have shown that the RF models can detect the dense cloud scenes and the large clear-sky areas with higher accuracy. However, the RF model has relatively low detection accuracy for broken clouds and thin clouds. This may be because the infrared radiation properties of these cloud FOVs and clear-sky FOVs are relatively similar.