Research On Global Surface Water Product Generation Based On FY-3/MERSI Data

Meteorological satellites play an important role in environmental monitoring.China's Fengyun series of meteorological satellites have served China's meteorological forecast and natural disaster monitoring for 51 years.A new generation of polar-orbiting satellites will be able to carry out round-the-clock data acquisition missions around the world.China's work on global surface water extraction needs to be further developed.It is the first to use 250 m resolution MERSI data to extract global surface water,which has important reference significance for the study of global change and the production of other series of products using this satellite data.This paper is based on FY-3D/MERSI data distributed worldwide.The research on basic processing methods and global data processing based on MERSI data has been carried out.On the basis of studying various methods of extracting land surface water information from remote sensing data,the research on water extraction from MERSI data is carried out based on adaptive single-Band threshold and adaptive normalized water index threshold.Four ensembled learning in machine learning of the water extraction experiment are studied emphatically.After the deep neural network was learned and the model was improved,the improved method was used to carry out the water extraction of the worldwide water quadrat.The water bodies in specific regions are extracted by using the global data of some time periods in 2020.The annual data were used to extract and analyze corresponding water bodies in the Yangtze River Basin in southern China.The research contents and experimental results of FY-3D/MERSI data in this paper are as follows:(1)According to different climate zones and land cover types on a global scale.A total of 139 samples were selected for the extraction of global surface water by combining the climate zone division and 30 m global land surface coverage FROM-GLC data.On this basis,data were screened for each sample according to multiple time periods,and water samples were collected.A total of 226,689 global water samples and 760,215 nonwater samples were collected.(2)On the basis of experiments with various water extraction methods.The corresponding model evaluation indexes were used to carry out the corresponding analysis work,and various models were used to extract the global surface water extraction quadrates.A DNN model suitable for water sample training is built on the basis of deep neural network learning research.According to the MERSI data water sample characteristics corresponding improvement and conversion,to carry out the global sample water extraction experiment.After results comparison and accuracy evaluation of various methods,it is considered that the Extre Trees has the highest accuracy.Deep neural network is the best method for water extraction in some small watershed.