Research On Cloud Detection Algorithm Based On Gaofen-5-DPC Data

The Directional Polarimetric Camera(DPC)on GF-5 satellite can observe the earth continuously with multiband,multiangle and high spatial resolution,and its data are of great significance to the study of atmospheric radiation transmission and cloud climate feedback in China.At present,there is no cloud recognition method established for DPC load.However,the traditional cloud recognition algorithm such as MODIS cloud recognition algorithm requires a wide range of band and the main detection band is from near infrared to thermal infrared,so DPC band cannot meet the requirements of this algorithm.The cloud recognition algorithm used by POLDER,which is similar to DPC,can be transplanted to DPC to a large extent.However,this algorithm has the following two problems: first,the recognition of snow and ice in polar regions and the selective recognition or non-recognition of solar flares in the ocean;Secondly,due to the error between the field of view of the instrument and the actual cloud distribution,it may be misjudged in the broken cloud and small cumulus pixel with small optical thickness.Therefore,this paper carries out the following research on the above issues:(1)Aiming at the traditional threshold method: based on the POLDER cloud recognition algorithm,this study developed a cloud detection algorithm suitable for DPC based on DPC multiband reflectance,polarization reflectance,apparent pressure and other information.The algorithm is divided into three parts.First,for cloud types with different heights,such as cirrus cloud and stratocumulus cloud,the algorithm introduces the apparent pressure test to increase the constraint conditions in the cloud recognition process.Secondly,cloud and clear sky identification are carried out in the land and ocean regions respectively through the different selection threshold bands of reflectance of different bands to surface objects.Finally,the reflectance of 865 nm polarization was used to identify the solar flare pixel reflected from seawater,which eliminated the misjudgment of solar flare when the reflectance threshold was used to identify the cloud pixel.(2)For the multiband and multiangle data of DPC,the vertical characteristic mask data of high-precision cloud-aerosol LIDAR and Infrared Iathfinder satellite observation(CALIPSO)were used as training samples in this study,and the cloud recognition results were obtained through the random forest method.(3)Compared with the traditional threshold method,the machine learning method has better universality,which can avoid the errors caused by many subjective factors in the physical method.However,due to the limitation of data samples,it is easy for the detection accuracy to decrease in the case of lower training samples.Therefore,this study combines these two algorithms to propose an optimal cloud recognition method.