Research On Image Fusion Of Gaofeng-2 And Sentine1-2 And Ground Object Classification

After more than half a century of development,the remote sensing technology system has gradually matured.In order to meet the increasing production or research needs,many types of sensors have been derived.Among them,the types of remote sensing images obtained by optical sensors are rich and widely used.Remote sensing has become a more scientific and effective means of modern collection of land cover information because of its strong data comprehensiveness and repeated periodic observation.At present,for optical imaging,due to the constraints of the internal optical system of the sensor,a single original data can not have high spatial resolution and high spectral resolution at the same time.When only using a single data source can not meet the demand,multi-source data collaboration can be considered.Among them,using image fusion technology to break through the limitations of the sensor on spatial and spectral resolution is an effective multi-source data collaboration method.Gaofen-2(GF-2)is the first civil optical satellite with sub meter resolution in China.Its panchromatic and multispectral data have better spatial resolution.For the low spectral resolution of Gaofen-2multispectral data(GF-2 MSS),Sentinel-2 MSI data can make up for it.Sentinel-2 MSI images have higher spectral resolution.The two are fused at the pixel level to achieve complementary advantage information.It is of great significance to the research of remote sensing image pixel level fusion and land cover classification.This thesis presents a fusion method combining band regression mapping(BRM)and highpass filter(HPF)for GF-2 MSS and Sentinel-2 MSI data sources,and uses this method to realize super-resolution(SR)reconstruction of Sentinel-2 MSI images and fusion of GF-2 MSS and Sentinel-2 MSI images.After the fusion processing,select the region with rich land types and complex scenes from the fused image as the study area for land cover classification.The conclusions of each part can be summarized as follows:(1)The resampling effect of the same interpolation method is basically the same for the experimental areas with different wave bands and different surface coverage types.The resampling performance of different interpolation methods is mainly affected by the resampling type(up sampling or down sampling)and the image resolution ratio.(2)Through the qualitative and quantitative evaluation of the fusion results,the feasibility of applying this fusion method to the two processes of Sentinel-2 MSI image super-resolution reconstruction and the fusion of GF-2 MSS and Sentinel-2 MSI images is confirmed.Among them,the results of HPF fusion using the predicted images after training the global regression model with four high bands are better.(3)This thesis compares and analyzes the classification performance of fused images,GF-2 MSS images,Sentinel-2 MSI images,and the classification performance of k-nearest neighbor(KNN)and random forest(RF)classifiers.According to the accuracy evaluation of experimental results,it is known that the classification accuracy of fused images is the highest when using the same training samples and classifiers.When using the k-nearest neighbor classifier,The overall accuracy is 2.68% higher than that of GF-2 MSS image,1.17% higher than that of Sentinel-2MSI image,and the Kappa coefficient is 0.06 higher than that of GF-2 MSS image and 0.03 higher than that of Sentinel-2 MSI image;When the random forest classifier is used,the overall accuracy is 12.17% higher than that of GF-2 MSS image,7.03% higher than that of Sentinel-2MSI image,and the Kappa coefficient is 0.11 higher than that of GF-2 MSS image and 0.08 higher than that of Sentinel-2 MSI image.In this thesis,the vegetation index(VI)is constructed based on the spectral information of the fused image.The performance of the VI is proved by comparative experiments.Finally,the k-nearest neighbor classifier with better performance is selected to classify the fused image object-oriented combined with a variety of features.The overall accuracy is 93.16% and the Kappa coefficient is 0.85.