Study On The Spatio-Temporal Fusion Algorithms

Multi-platform remote sensing satellites provide a large number of high spatial and temporal resolution remote sensing data,which provides an important data guarantee for accurate extraction of land cover and dynamic change information.In order to meet the needs of some applications,dense remote sensing data with high spatial resolution are needed.However,due to the limitations of hardware technology and budget,remote sensing images have the phenomenon that spatial and temporal resolution can not be obtained at the same time.There is complementarity between different sensor image information.Spatio-temporal fusion technology can generate remote sensing images with both high spatial and temporal resolution without changing the existing observation conditions,thus realizing dynamic monitoring of the surface at higher spatial and temporal resolution.Landsat data is a common data source in spatio-temporal fusion technology,but the ETM+ sensor carried on Landsat-7 has malfunction,which results in stripe missing of the acquired data,and greatly limits the availability of Landsat-7 data.A new object-oriented Landsat-7 SLC-off image interpolation repair algorithm is proposed to fill gaps.Spatio-temporal fusion technology still faces some challenges.On the one hand,due to the complexity of surface environment changes and the limitations of high-quality high-resolution remote sensing image acquisition,it is difficult to achieve the desired fusion accuracy when the complex changes of surface coverage occur.This paper proposes an improved spatio-temporal fusion algorithm based on the traditional spatio-temporal fusion research framework.On the other hand,many pixel-based spatio-temporal fusion algorithms are complex,and it takes a lot of time to process large quantities of remote sensing data.In this paper,deep learning algorithm is introduced to train appropriate spatio-temporal fusion model to improve the robustness and timeliness of the algorithm.This paper takes the spatio-temporal fusion of Landsat and MODIS remote sensing images as the research object.The specific research contents of this paper are as follows:(1)A new object-oriented Landsat-7 SLC-off image interpolation repair algorithm is proposed to restore the missing data caused by Landsat-7 ETM+ malfunction.The auxiliary image is firstly segmented at multi-scale to obtain the segmented objects,then the segmented objects are spatially overlaid with the image to be restored,the effective pixels in segmented objects are selected to calculate the value of missing pixels,and the local least squares regression parameters between the two images are obtained.The parameters are applied to the preliminary results to calculate the ultimate value of missing pixel.Experiments show that the algorithm can repair the spatial details of missing area,and has a certain practicability.(2)In view of the seasonal change of land surface and the change of land cover type,this paper proposes a spatio-temporal fusion algorithm based on linear interpolation model and neighborhood information.Firstly,a linear interpolation model is introduced to fuse the input high-resolution data and the low-resolution data of the predicted date.After obtaining the initial fusion results,in order to enhance the robustness of the algorithm,the neighborhood information is introduced to process the initial fusion results by spatial filtering,and the final fusion results are obtained.The experimental results show that the proposed algorithm can generate fused images with higher accuracy when seasonal changes occur,and can capture the spatial characteristics of terrain objects better when surface cover changes occur.(3)Considering the application requirement of remote sensing images with large scale and long time series and high spatial resolution,a spatio-temporal fusion algorithm based on deep learning is proposed.Firstly,the linear interpolation model is used for preliminary fusion,and then the residual dense network is introduced to reconstruct the preliminary fusion results.After the training phase is completed,the final fusion results are obtained by testing the input data.The experimental results show that the proposed algorithm has better prediction accuracy in areas with phenological changes.For data sets with land cover change as the main type,the algorithm still has better ability to predict the spatial structure of images.