| Subtropical plantation have become a popular direction for remote sensing research due to their complex structure and diverse tree species.Although a large number of remote sensing platforms have been put into use in recent years,a large number of data sources with different spatial and spectral resolutions can be used for forest remote sensing monitoring,but for different data sources,resolutions,and categorical variables,forest observations at different scales There are still few systematic comparative studies.In this study,the peak forest farm in Guangxi,southern China was used as the research area,using Zuyuan-3(ZY-3),Landsat-8,and Sentinel-2 images to improve its spectral and spatial resolution through data fusion;at the same time,the differences before and after fusion were extracted The spectral information and texture information of the data sources,combined with the digital surface model(DSM)and its derived information,construct three variable combination methods,and combine the actual situation of the study area to establish two fine and rough classification systems,using random forest algorithm Data optimization and classification,exploring the impact of different combinations of variables on the identification of subtropical forests under different spectral resolutions and spatial resolutions.the result shows:1.The data with a spatial resolution of 2m after fusion is classified based on two sets of fine and coarse classification system using the combination of spectrum,texture and terrain variables,can achieve the highest overall accuracy(83.48%,88.89% respectively).2.In the case of a spatial resolution of 2m,based on the coarse and fine classification systems using spectral variables for classification,the overall classification accuracy after data fusion is 72.07% and79.32% respectively,which is an increase of 14.2 compared to the pre-fusion %,14.66%.The improvement of spectral resolution can improve the classification accuracy of single tree species to different degrees.Among them,the classification accuracy of bamboo forest after fusion is improved by 44.35% compared with that before fusion;eucalyptus,coniferous forest,other broad-leaved forest,non-forest land can improve the classification accuracy of 10.64%-19.90%;the classification accuracy of shrubs is not obvious,only Increase by 2%.3.When the spatial resolution is 6m,data fusion can improve the classification accuracy of the image.Compared with the use of 6m ZY-3 multi-spectrum and Sentinel-2 before and after fusion,based on the classification results of spectral variables only,the improvement of spectral resolution can improve the classification accuracy by 11.13% and 3.71%,respectively.For single tree species,the improvement of spectral resolution can improve the classification accuracy of 4.66%-44.35%,but when the spatial resolution of the multispectral image before fusion is too low,it will reduce the classification accuracy of banded land types,such as shrubs.4.On the 2-30 m spatial resolution gradient,before data fusion,only the spectrum is used,and the spectrum and texture variables are used for classification.The overall image classification accuracy shows a trend of first rising and then falling,and Sentinel-2 data classification accuracy is the highest(68.21%,69.44%);use the spectrum,texture and terrain variables for classification,6 meters ZY-3 has the highest classification accuracy(74.19%).The data after data fusion is classified based on the spectrum.The 6-meter ZY-3 and Sentinel-2 fusion data has the highest classification accuracy;when using the combination of spectrum and texture variables,or the classification of spectrum,texture and terrain variables,the 2-meter ZY-3 The data fused with Sentinel-2 achieves the highest classification accuracy.5.For variables,the addition of texture variables can improve the classification accuracy of the image to varying degrees,and the addition of texture variables improves the classification accuracy of the image as the spatial resolution decreases.In the selection of texture variables,texture variables extracted based on large and small windows need to be used simultaneously.The most important texture variable for tree species classification in this study is Correlation(Cor).The addition of terrain factors can significantly improve the classification accuracy. |
