Precipitation Cloud Cluster Identification And Precipitation Intensity Level Estimation Based On FY-4A Satellite Data Based On Deep Learning

As an important part of the cycle of water resources on the earth,precipitation is significant for studying global climate change and monitoring natural disasters.Tibet is located on the Qinghai-Tibet Plateau,and its precipitation intensity level estimation work is affected by the scarcity of ground observation sites and traditional methods based on the physical characteristics of clouds are susceptible to the effects of bright ground and complex surface.Feng Yun-4A(FY-4A)satellite was used as the main data source,deep learning algorithm is used to study the precipitation cloud cluster identification and precipitation intensity level estimation algorithm suitable for the Tibet Plateau,and accurately estimate the precipitation intensity level within half an hour.A precipitation cloud cluster recognition algorithm based on improved DeepLab v3 is proposed,which uses atrous convolution to construct a multi-scale sampling module,and adds attention mechanism to extract deep high-dimensional features,optimized the upsampling method,effectively solve the imbalance of positive and negative samples,small precipitation cloud cluster feature extraction is insufficient and cloud cluster contour features are easy to be missed.Comparing the proposed method with the original DeepLab v3 and other models on the validation set,the experimental result show that the proposed algorithm had better segmentation and generalization performance,precipitation cloud cluster recognition results are more accurate,and Mean Intersection over Union(MIoU)reaches 0.95,which is 15.54% higher than the original DeepLab v3.On small targets and unbalanced datasets,this method can more accurately segment precipitation cloud clusters.A precipitation intensity level estimation model: Small Wisely Network(SW-Net)based on depthwise separable convolution is proposed,which greatly reduces the number of parameters while ensuring the performance of the algorithm.First,ablation experiment was used to determine the importance of different channels and channel differences for precipitation intensity level estimation,and then topographic data is added as auxiliary input data.Finally,it is compared with the existing high-performance precipitation estimation model Precipitation Estimation from Remotely Sensed Information Using Artificial Neural Networks–Convolutional Neural Network(PERSIANN-CNN)and U-Net.The results show that SW-Net provides a more accurate estimation of precipitation intensity,and the accuracy of the SW-Net model in various accuracy indicators has been improved,especially MIoU is higher than that of PERSIANN-CNN(U-Net)7.42(4.64)percentage points,indicating that the proposed model has better feature extraction capabilities.SW-Net has the lowest error,and the loss value and False Alarm Ratio(FAR)are respectively reduced by 23.89%(41.58%)and 18.27%(30.00%)compared with PERSIANN-CNN(U-Net),indicating that the proposed model has higher accuracy;In order to analyze the applicability of SW-Net,it has been applied to the North China Plain and the middle and lower reaches of the Yangtze River and achieves good experimental results.Compared with PERSIANN-CNN(U-Net),it also has better performance,MIoU has increased by more than 5 percentage points,and FAR has been drastically reduced by more than49%.It indicated that SW-Net was not limited by topographical factors,it is also suitable for plains and plateau areas.