Cloud Detection Algorithm Research And Prototype System Design Based On FY4 Data

Cloud,as an important part of the earth's water cycle,is of great significance for studying global climate change and monitoring natural disasters.Using meteorological satellite data to analyze cloud changes has become a major way.As the latest generation of stationary meteorological satellites in China,FY-4 satellite's high temporal resolution and high spectral resolution performance advantages are of great help to carry out cloud change analysis.Tibet is located on the Qinghai-Tibet Plateau.Its cloud detection work has a lot of problems such as the impact of snow and complex terrain.Effective cloud detection in Tibet is of great significance for monitoring precipitation in Tibet and it can reduce the impact of natural disasters.In this paper,the Tibet region is used as the research area,and the FY-4A satellite is used as the main data source to study the cloud detection in Tibetan Plateau.Based on this,an automated cloud detection system is developed in combination with modern software programming technology;The system can analyze the change of cloud cover in Tibet in 2019.The main research contents of this paper are as follows:(1)Try to complete the pre-processing of FY-4A /AGRI 4km resolution L1 data and Hamwari-8 satellite cloud product data to provide a scientific and reliable data source for cloud detection algorithms;(2)A multi-temporal multi-channel cloud detection algorithm is proposed.The algorithm makes full use of the characteristics of FY-4A's high temporal resolution with the different ranges of cloud and snow brightness temperature changes in continuous time to detect cloud.Compared with the FY-4A cloud detection product with the traditional single-phase cloud detection method,the accuracy of the multi-phase method is 90.4%,the false detection rate is 7.2%,and the missed detection rate is 5.6%,which is better than the other two methods.Based on this,the cloud phase recognition is further realized,and the recognition results are more consistent with the the CALIPSO data and GPM data;(3)An improved multiple feature fully convolutional network(MFFCN)is proposed for cloud detection.Based on the analysis of the native FCN network,the MFFCN's structure is simplified to speed up the training and prediction speed.At the same time,the performance advantage of FY-4A high spectral resolution is used,and MFFCN is designed to extract clouds with different spectral channels.The experimental result shows that the improved model is better for cloud detail detection and it is good at thin and broken cloud detection and the overall detection accuracy is 91.2%;compared with single-phase cloud detection method,multi-phase cloud method and cloud product data,the MFFCN has the highest detection accuracy and fast detection speed;(4)In order to complete the processing of massive remote sensing data,the Java and Python was used to design and develop the cloud detection system in Tibet;With the use of the system,the cloud detection work in Tibet throughout the year of 2019 is completed,The cloud cover was analyzed from two perspectives: cloud cover change and cloud phase change.