Forest Tree Species Classification Based On Multi-source Remote Sensing Data

Forests,as the mainstay of terrestrial ecosystems,play a crucial role not only in maintaining biodiversity,regulating global climate,maintaining ecological balance,and global carbon cycling,but also in providing irreplaceable benefits for human society,economy,and environment through permanent and sustainable use.To scientifically and reasonably manage forest resources,it is essential to understand and grasp information such as the types,quantities,and spatial distribution of forest resources.Forest species automatic recognition based on satellite remote sensing images can quickly obtain this information.Therefore,it is essential to establish a forest classification system applicable to large areas using economic satellite remote sensing data and other auxiliary data,and then obtain a high spatial resolution forest species classification map.This article focuses on the research of forest tree species classification in Liuzhou City and Luzhai County,Guangxi.Firstly,based on multi-spectral remote sensing images,a multifeature combination scheme,feature factor extraction,and different classification models were constructed.The impact of different features and machine learning models on tree species classification accuracy was evaluated,and suitable auxiliary data feature factors for tree species classification were determined.Secondly,based on hyperspectral remote sensing images,different hyperspectral dimensionality reduction methods were used to reduce the hyperspectral bands,and different ensemble learning models were used for forest tree species classification.The performance of different dimensionality reduction methods and ensemble learning methods were evaluated.Finally,different methods are used to combine the best classification results of hyperspectral data sets,multispectral images and panchromatic images to improve the image resolution.The forest tree species classification was carried out by fusing with suitable auxiliary data features,and the spatial distribution map of forest tree species in the best research area was generated through the comparison of different research schemes.The main conclusions are as follows:(1)The combination of Sentinel-2 imagery and auxiliary data can improve the classification accuracy of forest tree species,and the feature selection-based model had the highest classification accuracy among 16 feature combination models,the highest overall accuracy achieved was 82.69%.The spectral reflectance and spectral indices extracted from Sentinel-2 imagery can be used for forest tree species classification,but the value of texture features is limited and may even be negative.Auxiliary data,especially terrain features,UV aerosol index,phenology features,NO2 concentration features,terrain diversity features,precipitation features,temperature features,and multi-scale terrain position indices,play an important role in improving the accuracy of forest tree species classification.The RF,GTB,SVM,and CART algorithms were used to classify forest tree species based on Sentinel-2imagery,and the RF algorithm had the highest classification accuracy,with an overall accuracy of 82.69% and a kappa coefficient of 0.80.The overall accuracy was 0.14%,11.02%,and 11.70% higher than that of GTB,SVM,and CART,respectively.(2)By applying different dimensionality reduction methods to reduce the dimensionality of hyperspectral imagery,and combining them with ensemble learning algorithms,tree species classification was performed using various schemes,achieving a maximum overall accuracy of 84%.Four methods,PCA,ICA,BS-Net-FC,and BS-Net-CONV,were used to perform feature dimensionality reduction on the ZY1-02 D hyperspectral image to obtain corresponding datasets.The data set obtained by the BS-Net-CONV method had a higher overall classification accuracy and more suitable bands for tree species classification.The data sets obtained by the PCA,ICA,and BS-Net-FC methods also achieved a classification accuracy of over 80%,indicating that these methods are feasible for feature dimensionality reduction of domestic satellite hyperspectral data.In addition,four ensemble algorithms,Bagging,Boosting,Stacking,and Voting,were used to classify the dimensionality-reduced datasets for forest tree species classification.The overall classification accuracy of the Voting algorithm was the highest,with the highest overall accuracy of 84%.Compared with the highest overall accuracy,the Voting algorithm was 1%,2%,and 2% higher than Bagging,Boosting,and Stacking,respectively.(3)By utilizing various image fusion methods to enhance image resolution and comparing different classification schemes for multi-source data fusion,the optimal tree species classification scheme achieved an overall accuracy of 85.93%.A comparative study of fusion methods(GS,NND,and PC)for ZY1-02 D optimal spectral bands hyperspectral image,multispectral image,and panchromatic sub-meter spatial resolution image fusion was conducted using image quality evaluation.The GS fusion method had better overall evaluation results of fusion image quality,with higher image clarity and better preservation of the original image features.A comparative study was conducted on different classification schemes for multispectral,optimal hyperspectral images,and their respective fusion images.The results showed that the scheme of hyperspectral and panchromatic GS fusion with auxiliary data had the highest overall classification accuracy,which was 85.93%,with a Kappa coefficient of 0.84.Compared with the overall accuracy before the addition of auxiliary data,the accuracy was increased by 5.55%,and the Kappa coefficient was increased by 0.08.