| Vegetation resources are one of the material bases on which human beings depend for survival,and provide important ecosystem service for human production and life.Full monitoring of vegetation resources is an important prerequisite for assisting decision-making and tracking management.Remote sensing technology has become an effective means of monitoring vegetation resources due to its advantages such as strong current situation,wide coverage and fast timeliness.However,because of the strong correlation between the reflectivity of different vegetation types,the number of types that can be identified and the classification accuracy still need to be improved during the process of vegetation classification using remote sensing images.Especially in southern China where the distribution of vegetation is fragmented,the types are multifarious and the weather is cloudy and rainy,the uncertainty of remote sensing observation and the complexity of fine classification of vegetation types are significantly increased,leading to problems of unclear effective spectral deploying and combination modes of vegetation identification,unclear suitable spatial scale of images in southern fragmented planting areas,difficulty in improving the classification accuracy and so on.With the rapid development of remote sensing data in the direction of high spatial resolution,high spectral resolution,high frequency resolution and multiplatform and the increasing progress of remote sensing intelligent-processing technology,remote sensing technology is expected to make more breakthroughs in the fine classification of vegetation in highly fragmented planting areas in southern China.In this paper,S185 unmanned aerial vehicle(UAV)hyperspectral images,GF-1,GF-6 and Sentinel-2 satellite images are taken as the main data sources,Xingbin District of Laibin City in Guangxi is taken as the study area,and the spatiotemporal scale characteristics of multi-source remote sensing images are fully utilized.From three levels of spectral dimension pixel scale analysis,small-scale classification combining hyperspectral and high spatial features,and medium & large-scale extraction combining multi-temporal and red edge features,the fine classification of vegetation types from point and small area to region has been carried out respectively,thus realizing traceability of spectral separability of fragmented vegetation in south China,verification of suitable spatial scale of images,and achievement of large-scale fine extraction,and promoting the application and technical improvement of satelliteground multi-source remote sensing images,especially autonomous satellite data in fine classification of fragmented vegetation in south China.(1)The response characteristics of multiple combination modes and transform spectra to vegetation type recognition were revealed.In order to solve the problem of unclear effective spectral deploying and combination modes of vegetation identification,from the perspective of pixel-scale spectral dimension,taking the pixel-scale reflectance spectrum of S185 UAV hyperspectral image as a data source,starting from the scale of basic imaging unit,the response characteristic analysis of vegetation type identification based on band combination and transformation was carried out,and a response analysis method of pixel-scale reflectance spectrum based on multi-mode analysis and transformation for vegetation type identification was proposed.From three levels of mode establishment,load simulation and spectrum transformation,the characteristic spectra and their combination modes beneficial to the identification of vegetation types thereof were formed,the spectral separability was improved and the following conclusions were obtained through analysis.Firstly,the classification effect of the selected 23 bands achieved the accuracy based on all 101 bands,effectively reducing data redundancy and improving processing efficiency.Secondly,through a detailed analysis of the vegetation identification results of each band combination mode,it has been found that compared with the green-peak and redvalley edge bands and the yellow band,the red edge band can further improve the vegetation identification capability on the basis of the blue,green,red and near infrared bands.Therefore,further simulation experiment shows that the spectral deploying of GF-6 satellite has application potential in the fine identification of vegetation types.Finally,the effectiveness of spectral transform features in improving the identification accuracy was verified.(2)A high-resolution multi-scale model of vegetation fine classification coupled with feature space and scale parameters was constructed.In order to solve the problems of unclear suitable spatial scale range of remote sensing images in southern fragmented planting areas,the spatial scale suitability of fragmented vegetation classification in south China has been studied by using multi-scale S185 UAV hyperspectral remote sensing images as data source from the perspective of smallscale classification combining hyperspectral and high spatial characteristics.Integrated the coupling relationship between scale parameter optimization and feature effectiveness measure,from 7 scales of centimeter-decimeter-meter images,a fine vegetation classification model of high-resolution multi-scale remote sensing images coupled with feature space and scale parameters was constructed,and the regularity of feature space distribution and classification accuracy of multi-scale images were revealed,providing new evidence for the validity and scale suitability of spatial-spectral joint features in centimeter-level resolution image.The model classification accuracies of multi-scale images were 84.3%-91.3%,which were 1.3%-4.2% higher than that of Random Forest(RF)classifier under the same model framework and the same conditions.Compared with classification results of Support Vector Machine(SVM)classifier based on all features,the accuracies were improved by 2.5%-8.3%.The results show that the best classification accuracy does not occur in the original image,but at an intermediate level of resolution,and appropriate spatial scale range is beneficial to the improvement of vegetation classification accuracy.The study also found that the appropriate feature parameters in different scales have changed.With the decrease of spatial resolution,the importance of vegetation index features and textural features shows the opposite trend.The appropriate segmentation scale gradually decreases,with the appropriate number of features being 30-40.(3)A vegetation type information extraction mode with multi-temporal and red edge features of multi-spectral remote sensing image was proposed.In order to solve the problems of difficulty in improving classification accuracy of vegetation in south China from the regional scale,the multi-temporal and red edge assemblage features of vegetation classification was analyzed based on GF-1,GF-6 and Sentinel-2 images from the perspective of medium & large-scale vegetation information extraction.On the basis of spectral and temporal feature capture,a vegetation type information extraction model of multi-spectral remote sensing image with multi-temporal and red edge features was constructed from a hierarchical three-layer structure(single-temporal phase-> multi-temporal phase-> red edge feature),revealing the close relationship between phase and red edge features.Compared with the vegetation classification accuracy based on the single phase,the classification accuracy of GF-1 multi-temporal images was improved by 3.77%-14.22%;Sentinel-2 multi-temporal classification accuracy was improved by 5.59%-19.07%.Compared with the classification results without red edge features,the red edge band/red edge index promoted the vegetation classification accuracy of Sentinel-2 single/multi-temporal images by 1.51%-6.84%,with the highest accuracy reaching 86.09%.The study found that GF-1 and Sentinel-2 images of April have the highest accuracy and little difference in identification results,which is more suitable for fine classification of fragmented vegetation in south China.The classification accuracies of images in different phases are consistent with the seasonal variation of vegetation types.Comparative analysis of GF-6 and Sentinel-2 shows that the similar bands of the two have strong similarity.In the range of 0.45?m-0.85?m,the vegetation identification abilities of GF-6 and Sentinel-2 are equivalent. |
PDF Full Text Request