Research On Vegetation Classification Method Based On Multi-source Remote Sensing Data

As an important part of the ecosystem,vegetation has the ecological regulation functions such as water conservation and air purification.Vegetation cover classification is a hot topic of current remote sensing application research.With the continuous update and progress of remote sensing technology,the collection and acquisition of multi-source remote sensing data are becoming more and more convenient.Traditional optical sensors are susceptible to the influence of atmosphere,cloud and precipitation weather,which affect the ground classification accuracy,while synthetic aperture radar can work all day and all weather.The effective coordination of optical satellite images and synthetic aperture radar images can improve the accuracy of vegetation classification,and the collaborative classification of multi-source remote sensing data is also one of the important directions of remote sensing application.This paper takes the Northeast Tiger and Leopard Forest Park as the study area and investigates the fine classification method of vegetation based on the remote sensing images of the same period of Sentinel I and Sentinel II with multi-source remote sensing combination data.Firstly,multi-scale segmentation of multi-source images is carried out by combining the ESP scale evaluation tool and the mean dissimilarity index;secondly,random forest screening and feature variable analysis are carried out based on the feature covariates of multi-source combined images;finally,different machine learning classifiers are used for object-oriented classification,and comparative analysis of classification methods is carried out to select a suitable classification algorithm.The main research results of this paper are as follows:(1)In order to select the best scale parameter for the combined multi-source remote sensing images,we used the ESP segmentation scale evaluation tool and performed the average heterogeneity calculation on the obtained multiple scale parameters.After analysing and comparing the calculated under-segmentation and over-segmentation results,we determined the optimal scale parameter to be 328.Our study shows that combining the ESP scale evaluation with the calculation of the mean heterogeneity index can improve the accuracy and efficiency of the optimal scale segmentation parameter selection.(2)We used the random forest method to optimally filter and reduce the dimensionality of feature information extracted from different bands of multi-source remote sensing data in order to reduce data redundancy and improve classification accuracy,and finally reduced the 89 feature variables optimally to 23.The feature variable analysis we conducted showed that these filtered features could clearly distinguish different features,and the inclusion of feature information from SAR images could enhance the separability of features.(3)We used an object-oriented approach based on combined multi-source remote sensing imagery to finely classify vegetation and compared the classification results of three classification algorithms.The results showed that the random forest algorithm and C5.0 decision tree algorithm achieved a classification accuracy of 91.33% and89.59% respectively,while the support vector machine algorithm achieved a classification accuracy of 58.96%.After qualitative and quantitative comparisons,we found that the random forest algorithm had better classification results and could distinguish between different vegetation cover types.