| Forest resource survey plays a crucial role in mapping the quantity,quality,and distribution of forest resources.Utilizing remote sensing technology,which involves non-contact detection techniques,significantly reduces the workload associated with external tasks involved in surveying and monitoring forest resources.Consequently,it offers a promising avenue for extracting forest land and forest tree information.In the realm of remote sensing applications in forestry,one of the primary areas of research revolves around selecting different types of remote sensing data sources based on factors such as spectral resolution,spatial resolution,temporal resolution,and coverage.This consideration is necessary to meet the research scale and forest property requirements for effective forest resource monitoring.Furthermore,the rapid advancement of machine learning algorithms in remote sensing data processing technology serves as a crucial pillar in supporting the extraction and spatial distribution of forest information.This technology is applicable across various scales,including air,sky,and ground,and it harnesses satellite remote sensing data,UAV remote sensing,and Li DAR data to accomplish the extraction of valuable forest information.In this research study,the Support Vector Machine(SVM)and Random Forest(RF)algorithms were employed to extract land use information.Landsat-8,Sentinel-2A,and GF-1remote sensing images served as the base data sources for SVM,RF,and respective analyses.The study focused on the Dulaer forestry located in Arxan City as the specific target area.Moreover,the research investigation concentrated on forest land,utilizing UAV multispectral remote sensing data to construct a tree species classification model.This model integrated spectral vegetation index and texture features as key inputs and underwent a rigorous evaluation of its accuracy.Furthermore,leveraging UAV Li DAR data,a Canopy Height Model(CHM)was established to extract tree height information.This approach facilitated the acquisition of precise vertical information regarding the forest’s structure.The main research findings are summarized as follows:(1)The SVM algorithm applied to Sentinel-2A achieved an overall classification accuracy of 86.59% with a Kappa coefficient of 0.807.Similarly,the RF algorithm yielded an overall classification accuracy of 87.39% with a Kappa coefficient of 0.816.For the GF-1dataset,the SVM algorithm attained an overall classification accuracy of 79.21% with a Kappa coefficient of 0.714,while the RF algorithm achieved an overall classification accuracy of80.71% with a Kappa coefficient of 0.733.Regarding Landsat-8,the SVM algorithm resulted in an overall classification accuracy of 76.04% with a Kappa coefficient of 0.689,while the RF algorithm achieved an overall classification accuracy of 79.76% with a Kappa coefficient of0.737.Significantly,the RF algorithm demonstrated the highest accuracy for the automatic extraction of land classes in the study area when using Sentinel-2A imagery.(2)Utilizing UAV multispectral data for tree species classification,a comprehensive analysis was conducted on 59 spectral vegetation indices and 91 spectral vegetation indices combined with texture features.Through the implementation of ANOVA and Successive Projections Algorithm(SPA),18 indices and 23 indices were respectively selected to construct classification feature sets.The SVM algorithm was then applied to these feature sets for tree species classification.Remarkably,the SVM algorithm achieved an overall accuracy of 96.72%and a Kappa coefficient of 0.957 when utilizing the spectral index feature set.Furthermore,the most influential features within this set were identified as SI1*,DVIreg,SI3*,OSAVIreg,INT*,and DVI* based on their importance rankings.Similarly,for the spectral index + texture feature set,the top 6 sensitive features were determined as SI1*,DVIreg,9*9MEA,OSAVIreg,SI2 re,SI2*.Four distinct tree species classification models employing SVM and RF algorithms were developed using these selected features.The SVM algorithm,based on the spectral index feature set,achieved an impressive overall accuracy of 96.72% and a Kappa coefficient of 0.956.In contrast,the RF algorithm achieved an overall accuracy of 75.41% and a Kappa coefficient of 0.709 for tree species classification.Moreover,employing the spectral vegetation index + texture feature set,the SVM algorithm attained an overall accuracy of 83.61%and a Kappa coefficient of 0.8,while the RF algorithm achieved an overall accuracy of 72.13%and a Kappa coefficient of 0.675.These findings highlight the significance of selecting appropriate algorithms and a small yet essential set of features to ensure accurate differentiation of different tree species and achieve high classification accuracy.(3)Using UAV Li DAR data,tree vertices were extracted from CHM images at resolutions of 0.1m,0.2m,0.3m,and 0.4m,employing a single wood recognition algorithm.To enhance accuracy,mean and median filtering were applied with different window sizes.The optimal result for extracting reference tree vertices was achieved by employing a local maximum visual decoding method with 3×3 windows on the CHM0.1m.Additionally,tree heights were extracted utilizing CHM images with varying resolutions.Regression analysis was conducted to establish the relationship between the extracted tree heights and the corresponding measured tree heights.The best fit was observed between the extracted tree heights derived from the CHM0.4m and the measured tree heights,resulting in an R-squared value of 0.97(P < 0.01)and an RMSE of 0.8m. |
