Research On Global Sea Surface Salinity Inversion Method Based On MODIS Data

In recent years,satellite observation technology has been developping rapidly,leading to improved the global ocean monitoring capabilities.Global ocean surface salinity data is an important parameter for studying scientific issues such as ocean circulation,global water cycle,and climate change,but for satellite observations of ocean salinity,the low spatial resolution(30 to 100 km)and low temporal resolution(3 days or more)limit the applicability of satellite observations of ocean surface salinity products.Therefore,in order to be able to quickly invert the global ocean surface salinity data and improve the spatial resolution of sea salt products,a deep learning method is used to invert the global sea surface salinity.The main work and conclusions are as summarized a follows:1.Using the improved convolutional neural network regression method as an inversion model,the functional relationship between MODIS seawater reflectance products and seawater temperature products and global sea surface salinity data is established.In the experiment,the influence of different convolution methods,different pooling methods and different model complexity on the inversion accuracy was analyzed in detail.Experiments show that different convolution and pooling methods have little effect on the inversion accuracy.When the model complexity is 11 layers,the inversion accuracy is high.The decision coefficient of the optimal model is0.94198,and the root mean square error is 0.06967.From the experimental results,the improved convolutional neural network inversion model used in this paper has good fitting accuracy and can quickly and accurately invert the global sea surface salinity.2.In order to improve the spatial resolution of sea salt products,remote sensing downscaling technology is added to the existing inversion model.In this paper,a deep convolutional network(DCN)model is designed to combine high spatial resolution sea surface reflectance data,seawater temperature data and low spatial resolution SMOS sea surface salinity data to improve Spatial resolution of sea surface salinity data.The DCN model in this paper consists of two basic modules:(1)feature enhancement module and(2)downsampling module.In the research,this papercompares the effect of the improved CNN,U-net,Resnet,and FCNN as feature enhancement modules on the DCN model inversion accuracy,and the effect of using different scale input data on the inversion accuracy.Experiments show that when Resnet and U-net networks are used as feature enhancement modules,the DCN inversion model has the highest accuracy.The decision coefficient of the optimal model is 0.94228,and the root mean square error is 0.02191.The results show that the DCN model can improve the spatial resolution of the product while inversion.