Study Of Object Classification Based On Multispectral Images Of UAV

Acquisition and classification of remote sensing image is the basis and key technology in the process of remote sensing monitoring.Different from the traditional aerospace remote sensing of low spatial resolution,temporal resolution and vulnerable to the effects of atmospheric environment,rapid development of unmanned aerial vehicle(UAV)and lightweight sensors brings out the possibility of low altitude remote sensing and high spatial resolution,high spectral resolution image.At the same time multi-spectral images tend to produce "same-spectrum foreign body","foreign body with the spectrum" and so on;which has brought difficulties to follow-up's treatment.Higher spectral resolution also increases the correlation between adjacent bands,and consequently a large amount of information redundancy,which not only brings computational complexity but also increases the time complexity.Therefore,it is a difficult problem in the application research to reduce the dimension processing of multi(high)spectral remote sensing data.Oriented low-altitude multi-spectral features classification,and then based on the best band index and the spectral characteristics,texture feature of image to choose the best combination of bands.Finally,the support vector machine and the least squares support vector machine(SVM)are used to build the classification model for classification and comparison experiments.The main work and related research results are as follows:(1)The remote sensing image acquisition platform of UAV was built by large-scale fixed wing unmanned aerial vehicle,which equipped with light multi-spectral camera,and then the remote sensing images of UAV with 22.6cm GSD and 12 bands were obtained.Then the original image was registered and fused at the characteristic level to obtain the orthographic image of the study area by Pix4 D Mapper.(2)Aiming at the characteristics of high spatial resolution and high inter-band correlation of UAV multi-spectral image data,synthesize spectral information such as vegetation and water related index,texture feature information acquired by PCA and GLCM,the original bands which was filtered by the best band index method to obtain the best band combination for feature classification.(3)The unsupervised and supervised classification methods were designed to classify the objects in the study area.Compared to the original band combination,the accuracy of Iso Data classification of the 1,6,11,NDVI,NDWI and Mean bands in the study area A increased from 83.57% to 89.80% and the SVM classification accuracy was increased from 95.58% to 99.76%.Experiment results confirm that the best bands combination not only has more band information and the inter-bands correlation coefficient is lower,but also reflects the spectral information and texture information of objects,so it can be selected as the best bands combination for the Micro MCA12 Snap.(4)According to the best bands combination from the experiment,SVM and LSSVM experiments were carried out on the study area by using particle swarm optimization and grid search algorithm for parameter optimization and cross validation.The LSSVM with PSO classification model is obtained by parameter optimization of particle swarm optimization.Compared with the SVM particle swarm optimization,the classification accuracy improved from 97.833% to 99.854%;Compared with the LSSVM grid search,the classification accuracy improved from 99.762% to 99.854%.At the same time,LSSVM particle swarm optimization improves the classification speed to a certain extent,and is an ideal classification model for the classification of the best bands combination in this paper.