Research Of Spatiotemporal Fusion Technology Of Land Cover Remote Sensing Data

With the development of remote sensing,the application of remote sensing data is more and more extensive.However,due to the limitation of sensor design,people can’t get remote sensing data with high spatial and temporal resolution.In addition,clouds and other atmospheric conditions further limit research in some areas.Therefore,a single data source usually can’t satisfy the needs of research.In order to solve this problem,many spatiotemporal data fusion methods have been developed,which aim to integrate two or more kinds of remote sensing data and generate data with high temporal and spatial resolution.In recent years,with the development of artificial intelligence,some deep learning algorithms have begun to be applied in spatiotemporal fusion tasks and have high potential in application.This thesis provides new solutions for the spatiotemporal fusion tasks with the help of theories in deep learning.The main work of this thesis are as follows.（1）Aiming at the problems that the existing learning-based spatiotemporal fusion methods cannot extract the spatiotemporal information between images very well and rely on weighting functions for image fusion,this thesis proposes a spatiotemporal fusion method based on attention mechanism.This method uses two pairs of high temporal resolution but low spatial resolution and high spatial resolution but low temporal resolution images on the reference date,and one high temporal resolution but low spatial resolution image on the predicted date,to fuse a high spatial resolution image on the predicted date.The main structural of this method is a two-stream network with multiscale residual attention blocks which combine multi-scale feature extraction mechanism and channel attention mechanism.In the process of image fusion,an image fusion module is proposed,which is based on the channel attention mechanism,to improve the accuracy of the fusion result through the learning ability of the network itself.In the training process,a compound loss function is used to improve the learning effect of the network.Through experimental comparison,this method improves the effect of spatiotemporal image fusion.（2）A progressive spatiotemporal fusion method based on attention mechanism is proposed to overcome the shortcomings of pre-upsampling methods.This method adopts a progressive upsampling method to gradually increase the resolution of the predicted image.On the other hand,multi-scale residual attention blocks is used to enhance the ability of learning and prediction accuracy of the network,and more spatial details are preserved by adding skip connections.A compound loss function is also used to optimize the network better.Through experimental comparison,this method further improves the effect of spatiotemporal image fusion.