Research On Larch Forest Recognition Based On Sentinel Data

As the main plantation species in China,larch has played a great role in greening the country,improving the ecological environment and increasing the main by-products of forestry.In recent years,the spatial distribution of larch plantations is still unclear because the area of larch plantations is increasing year by year,and the large area planting has caused serious problems such as single species,weak resistance to environmental changes and pests and diseases.Therefore,how to quickly obtain the spatial location distribution of larch plantation forests has become a hot issue in larch plantation forest cultivation research.Along with the rapid development of remote sensing technology,the use of remote sensing data to quickly and accurately obtain the spatial distribution of larch plantation forests is of great significance for the cultivation,rational utilization and planning management of larch plantation forests.In this paper,the remote sensing images of Sentinel-1 and Sentinel-2 in typical periods were selected based on the phenological and seasonal characteristics of larch plantations in Seyhanba Forest in Hebei Province.Based on the optical data,radar data,and the combination of radar and optical data,multiple features of larch plantations were extracted and selected.The spatial distribution of larch plantation forests was extracted by Random Forest(RF),Maximum Likelihood Classification(MLC)and Support Vector Machine(SVM)classification algorithms,and the performance of the classification algorithms was evaluated.The performance of the classification algorithms was evaluated.In order to obtain the best classification time phase,feature combination and algorithm for extracting larch plantation,and to provide a method reference for quickly obtaining the spatial location distribution of larch plantation.The main results of the study are as follows.(1)Single-temporal phase,multi-temporal phase and different methods of larch plantation forest identification based on Sentinel-1.The results show that among the temporal phases,the best classification phase is March,the highest accuracy of RF,MLC and SVM using VV and VH backward scattering coefficients are 63.82%,60.28%and 62.54%,respectively,the Kappa coefficients are 0.59,0.55 and 0.57,and the larch extraction accuracy are 76.48%,72.52%and 74.09%;their classification accuracies combined with texture features improved by 3%-8%,indicating that texture features can improve the classification accuracy of larch plantation forests.Therefore,the best classification results were obtained in March using the VV and VH backscattering coefficient features of RF and texture features.Combining the multi-temporal feature combinations and feature importance selection,the classification accuracies of RF,MLC and SVM were 73.2%,65.5%and 67.41%,and the extraction accuracies were 83.39%,76.12%and 77.33%,respectively.Among them,RF has the highest classification accuracy and extraction accuracy.Compared with the feature combination of single temporal phase,the classification accuracy and extraction accuracy of the combined multi-temporal phase feature combination are higher.Therefore,the RF algorithm is most suitable for the classification of larch plantation forests in combination with the feature preference of multi-temporal phases.(2)Single-temporal phase,multi-temporal phase and different methods of larch plantation forest identification based on Sentinel-2.RF,MLC and SVM were used to extract larch plantation forests using the spectral features of Sentinel-2 images and vegetation indices.The results showed that SVM,RF and MLC had the highest classification accuracy of 90.9%,90.1%and 88.42%for larch plantation forest in March using only mono-temporal spectral features,and 93.11%,91.32%and 89.27%for extraction accuracy,respectively,indicating that spectral features have an important role in extracting larch plantation forest.Further combining spectral features and vegetation indices using RF,MLC and SVM,the highest classification accuracies were 89.33%,83.75%and 82.25%,and extraction accuracies were 92.02%,84.24%and 84.11%,respectively,which decreased 1.6%～6.4%in classification accuracy and 1.1%～7.1%in extraction accuracy relative to larch using only spectral features.This indicates that the inclusion of vegetation indices in the mono-temporal phase has no significant effect on the classification accuracy of larch plantations.Finally,the classification accuracies of RF,MLC and SVM were 94.82%,93.7%and 94.54%,respectively,and the extraction accuracies were 94.99%,92.92%and 94.48%,respectively,by combining the spectral features and vegetation indices of multi-temporal phases,among which RF had the best classification effect.Therefore,the classification strategy using the spectral and vegetation index features of the integrated temporal phase is the best.(3)Combining Sentinel-1 and Sentinel-2 with mono-temporal,multi-temporal and random forest feature factor importance assessment to establish feature library larch plantation forest identification.The results showed that the classification accuracy of the feature set built by combining each corresponding mono-temporal phase of Sentinel-1 and Sentinel-2 using RF was 88.4%,82%,88.63%and 88.15%,and the extraction accuracy was 91.88%,84.35%,89.34%and 89.82%,respectively.Among them,the classification accuracy was highest in June and the extraction accuracy was highest in March.Therefore,the most accurate extraction of larch plantation forest was achieved in March by combining spectral,vegetation index,VV.VH and texture features using RF classification strategy.In addition,the feature set constructed based on random forest feature importance assessment by combining the features of the four temporal phases of Sentinel-1 and Sentinel-2 had 91.25%classification accuracy using RF,92.41%extraction accuracy,and a Kappa coefficient of 0.88.Compared with the best classification strategy of single temporal phase,the classification of multiple temporal phases was better,indicating that the radar feature data combined with optical feature data has some potential for larch plantation extraction.