Reconstruction Of MODIS NDSI In Tibetan Plateau Based On GAN Neural Network

With the intensification of global warming in recent years,more and more researches focus on the field of remote sensing of snow cover.As the most widely distributed cryosphere in China,the Tibetan Plateau contains abundant snow resources,which have a profound impact on the climate system of my country and the world.MODIS NDSI snow product is widely used in snow research,but the product will have numerical gaps caused by cloud cover.How to accurately reconstruct MODIS NDSI numerical gap products has important research significance.At present,the existing research mainly uses the space-time interpolation algorithm or simple machine learning algorithm to reconstruct the vacant value area,which is relatively simple for the distribution characteristics simulation of NDSI continuous values,and there is still room for improvement in the accuracy index.In this study,based on Generative Adversarial Networks(GAN),the MODIS NDSI data reconstruction model of the Tibetan Plateau was constructed,and the Mean Absolute Error(Mean Absolute Error,MAE),Root Mean Squared Error(Root Mean Squared Error,RMSE),Positive Mean Error(PME),Negative Mean Error(NME),and Structural Similarity(SSIM)are used to evaluate the accuracy of the reconstruction results.Firstly,combined with the spatiotemporal information of MODIS NDSI products,the gaps in the original data were preliminarily filled,and the daily MODIS NDSI images with effectively reduced number of cloud pixels were obtained,which were used as the input data of the GAN neural network to construct a single-channel GAN neural network model.It is found that there are obvious overestimation and underestimation errors in the filling results of the single-channel GAN neural network,and there is still room for improvement in the structural similarity between the filling results and the true distribution of NDSI values.Next,this study constructed a three-channel GAN neural network model by further introducing SRTM elevation data and air temperature data,and found that compared with most machine learning filling methods in recent years,the three-channel GAN neural network can effectively alleviate the overestimation and underestimation errors.Clearly restore the spatial texture distribution characteristics and significantly improve the structural similarity.The main research content and results of this paper are as follows:(1)Use the space-time cube interpolation algorithm to preliminarily fill in the data gap pixels: combine the MODIS NDSI products of the two days before and after,construct a 3×3×3 space-time cube,and fill in the missing pixels.The filling accuracy is RMSE=6.496,and the residual cloud rate is 45.966% of the initial cloud amount.This data can provide important input data for subsequent GAN neural network training.(2)Construct a single-channel GAN neural network data reconstruction model:use the MODIS NDSI product interpolated by the space-time cube to make a series of slice data sets with a size of 128×128,and pack each 64 pieces into a batch as the input of the training network data.The VGG16 network framework is used to construct the discriminator of the GAN neural network,and the convolutional neural network is used to construct the generator of the GAN neural network.When the training reaches convergence,the reconstruction accuracy of the generator model on the verification set is RMSE=13.801,MAE=8.879,PME=8.138,NME=-27.574,and there are obvious overestimation and underestimation errors.(3)Constructing a three-channel GAN neural network data reconstruction model:MODIS NDSI products,SRTM elevation data and air temperature data are used as three-channel data after the space-time cube interpolation,and each channel makes a series of slice data with a size of 128×128,each 64 sets of three-channel slice data is packaged into a batch as the input data for training the network.The VGG16 network framework is used to construct the discriminator of the GAN neural network,and the convolutional neural network is used to construct the generator of the GAN neural network.When the training reaches convergence,the reconstruction accuracy of the generator model on the verification set is RMSE=9.125,MAE=6.811,PME=6.777,NME=-16.370,which is significantly improved compared with the reconstruction result of the single-channel GAN generator,and SSIM= 0.896,which proves that the three-channel GAN generator model has a good structural similarity recovery ability.At the same time,from the statistics and analysis results of the accuracy of the elevation and temperature partitions of the generated results,the generator of the three-channel GAN neural network constructed in this study has better performance in areas with elevations lower than 4000 m and temperatures higher than 0 °C.