Research On Land Use Classification Based On Sentinel-2 And MODIS Spatio-temporal Fusion

Using remote sensing images to classify land use plays a vital role in the rational planning and utilization of land resources.Remote sensing technology can quickly obtain large-scale land use information in a short period of time.With the continuous development of satellite technology in recent years,land use classification based on remote sensing images has become the mainstream classification method.Long-term high-resolution remote sensing images can better reflect the phenological information of ground objects,and have a good effect on improving the classification accuracy of ground objects.At the same time,it has the characteristics of high temporal resolution and high spatial resolution.The Sentinel-2 image with high spatial resolution is affected by cloudy rainy weather and atmospheric refraction,resulting in serious image loss and reduced classification accuracy.The high temporal resolution of MODIS images is limited by its own sensor,and the spatial resolution is too low,and the phenomenon of "same-spectrum foreign bodies" is prone to occur when it is used for land use classification.Therefore,how to use these two kinds of images to construct remote sensing images with high spatial and high temporal resolution is a focus of this research.In view of the complex topographical conditions and diverse structural forms in the Henan section of the Yellow River Basin,it is very important to study efficient and accurate land use classification methods for the acquisition of land use information.This paper takes Zhongmu County,Zhengzhou City,Henan Province as the research area,based on the fusion of Sentinel-2 time series remote sensing images,and explores a high-precision land use classification method through classification model optimization.The main work and conclusions are as follows:(1)Multi-source image fusion algorithm is preferred.Taking Sentinel-2 image and MODIS image as data sources,this paper conducts experiments on four pixel-level fusion algorithms,Brovey,FSDAF,ESTARFM,and G-S,and evaluates the accuracy of fusion images from both qualitative and quantitative aspects.The experimental results show that the correlation between the ESTARFM fusion image and the real image is the highest,and the correlation coefficient R of each band is 0.9525 and0.9761.algorithm.Overall fusion accuracy is sorted from high to low: ESTARF >FSDAF > Brovey > G-S.(2)A new fusion strategy-"alternative fusion" is proposed.When encountering continuous cloudy and rainy weather,the time span of the two Sentinel-2 images will be relatively large.At this time,ESTARFM is used for fusion,and the reliability of the results needs to be discussed.The FSDAF algorithm is introduced,and the FSDAF algorithm is used to fuse the different images of the two scenes at the same time.The experimental results show that the images fused by the "alternating fusion" method have higher correlation with the real Sentinel-2 images of the same period.(3)Land use classification algorithm optimization and classification.The HANTS time series harmonic analysis method is used to filter and reconstruct the time series NDVI curve,which improves the classification accuracy of vegetation.The classification experiments were carried out on support vector machines,artificial neural networks,and random forest algorithm without feature optimization,and the accuracy of the classification results was evaluated by using confusion matrix.The experimental results show that the support vector machine method has the highest accuracy and the fastest running speed,the overall classification accuracy is 94.0358%,and the Kappa coefficient is 0.9273,which is higher than the artificial neural network algorithm and the random forest algorithm without feature optimization.(4)Aiming at the Sentinel-2 time series images in 2019,the land use classification was carried out by using the support vector machine method,and the accuracy of the classification results was evaluated by using the confusion matrix.The results show that the overall classification accuracy is 90.5422%,and the Kappa coefficient is0.8849,which is lower than the fusion time series image classification results.The experimental results show that the long-time series remote sensing image dataset is constructed based on the spatio-temporal fusion algorithm,and the support vector machine method is used for land use classification,which can improve the classification accuracy to a certain extent.27 figures,15 tables,and 82 references.