Research On Cloud Detection And Its Motion Prediction Based On Deep Learning

Cloud plays an important part in radiation balance and the water vapor cycle.Therefore,cloud detection and movement prediction tasks are significant for weather prediction and disaster warning.However,traditional cloud detection algorithms perform far from satisfactory due to the fuzzy boundaries and complex textures of clouds.Although deep learning methods have shown superior performance in cloud detection and movement prediction,they are constrained by limited labels in ground-based cloud image data sets and the structure of the deep learning model.Aiming at the above problems,based on the deep learning method,this paper studies two aspects: cloud detection and cloud movement prediction.The relevant research contents and results are as follows:(1)A new ground-based cloud detection(GBCS)dataset,which contains 1742 accurately labeled images collected from Data Fountain Competition of weather identification and internet,was established for the cloud detection task.Then to evaluate how well deep learning models perform in cloud segmentation,12 state-of-the-art semantic segmentation networks were selected,among which DeepLabV3+ outperformed all others.Furthermore,since 1742 images are not enormous,a novel Transfer learning(TL)-DeepLabV3+ model was developed by TL.TL-DeepLabV3+ showed a high ability of cloud segmentation,scoring the Mean Intersectionover-Union(MIo U)of 91.05% on GBCS and further verified in the UTILITY data set and the Cirrus Cumulus Stratus Nimbus(CCSN)data set.Compared with traditional cloud detection methods and deep learning methods,the cloud detection ability and generalization ability of TL deeplabv3 + are greatly improved.(2)The Himawari-8 satellite data was used to make a spatiotemporal sequence dataset for the cloud movement prediction task.This paper proposes a new model PredGAN by combining the Generated Adversarial Network(GAN)with a state-of-the-art spatiotemporal sequence prediction model.The results show that the PredGAN performs better than other models.Moreover,especially for long-distance prediction,PredGAN can still maintain a high level,and there is no apparent blur in the predicted cloud images.Finally,PredGAN is used to test an example outside the time range,proving that it has strong generalization ability and can accurately predict the infrared cloud image after 200 minutes.The prediction time is improved compared with the traditional method,which is significant for nowcasting.