Classification Of Mangrove Species Using High-resolution Multi-source Remote Sensing Images

As a globally recognized ecosystem rich in carbon storage,mangroves can play multiple roles such as flood and tide resistance,coastal erosion prevention,pollutant filtration,seawater purification,carbon sink,ecological environment improvement,and biodiversity maintenance,providing irreplaceable social,economic,environmental and ecological services for human and coastal organisms.However,under the influence of the natural environment and human activities,mangrove forests are being continuously damaged.In order to reduce losses,protect and restore mangrove forests,fine monitoring of mangrove species distribution is needed to obtain accurate information and mapping,which helps to understand the process of ecological changes along the coastal zone and provides the theoretical basis for resource planning and investigation of mangrove forests and development deployment of ecological improvement.The special geographical environment of mangroves poses considerable challenges to species classification,making remote sensing technology a major tool for research and management.The research on the identification and classification of mangroves based on remote sensing images started early,but the accuracy and effect of species classification have been unsatisfactory.This paper takes the southern part of the mangrove reserve in Touyuan Village,Bamen Gulf,Qinglan Harbor,Wenchang City,Hainan Province as the research area,and tried to investigate the performance of the unmanned aerial vehicle(UAV)Rikola hyperspectral image,World View-2(WV-2)satellite-based multispectral image,and fused data from both in the species classification of mangroves.We used Recursive Feature Elimination-Random Forest(RFE-RF)algorithm to select vegetation spectral and texture features,and used Random Forest(RF)and Support Vector Machine(SVM)in machine learning to classify mangrove species.The advantages and shortcomings of different data sources and different classification methods were compared,and the driving factors affecting the classification accuracy were discussed.The main study contents and findings are as follows.(1)Vegetation index features of UAV hyperspectral images and RF algorithm for more effective species classification of mangroves.The research of UAV hyperspectral data mining and dimensionality reduction methods was carried out in terms of feature extraction and band selection,and the 15 vegetation index features and 1080 texture features extracted were used to downscale the hyperspectral data using RFE-RF method,and the five most important features were selected(H?MSAVI,H?SR890,H?NDVI800,H?OSAVI2,H?NDVI750,H?SR750),demonstrating that the spectral features of high spatial resolution and high spectral resolution images play a significantly larger role than texture features.The species classification of mangroves using RF and SVM algorithms for the preferred feature set of UAV images.Dual validation with sample points established based on UAV images and GPS waypoints showed that the RF algorithm performed the best,with an overall classification accuracy of 92.70%,an improvement of 20.21% compared to the SVM(RBF)algorithm,and a kappa coefficient of 0.91,an improvement of 0.23 compared to.In terms of classification accuracy of individual tree species,the best performance was still the RF algorithm,which achieved producer accuracy and user accuracy for all five tree species reached more than 87%,and the best separability was achieved for Sonneratia caseolaris and the worst for coconut palm in the five tree species categories.(2)The accuracy of mangrove species classification based on WV-2 satellite images was significantly lower than that based on UAV hyperspectral images,but principal component analysis of WV-2 images and extraction of texture factors helped to improve species classification accuracy.The problem of sample point mismatch between WV-2 images and UAV hyperspectral images was effectively mitigated by sample sampling thinning.Based on WV-2 images,the most core component features were integrated using principal component analysis,and the same extraction and dimensionality reduction process was used to preferentially select the most important five features in order as WV?PCA3?mean?7,WV?PCA2?contrast?7,WV?PCA3?sm?7,WV?B?pan and WV?EVI.It was found that the principal component analysis of WV-2 images helped to improve the species classification accuracy of mangroves,and the role status of texture features was high in WV-2 images.The accuracy of mangrove species classification using RF and SVM algorithms for the selected feature set of WV-2 images was lower than that of UAV hyperspectral images.(3)The mangrove species classification mapping accuracy can be further improved by fusing the vegetation index features of UAV hyperspectral images with the texture features of WV-2 images.Based on the ultra-high-dimensional data set consisting of a total of 1471 features from UAV hyperspectral images and WV-2 images,the five most important feature variables(WV?PCA3?mean?7,H?MSAVI,WV?PCA2?contrast?7,WV?B6?entropy?5,H?OSAVI2)were preferentially selected using RFE-RF,and it was found that as the spatial and spectral resolution of the images increased,the role of texture features in the species classification of mangroves in Qinglan Harbor was decreasing,while the role of vegetation index features was increasing.The species classification of mangroves was performed using a parameter-optimized machine learning algorithm,and the accuracy of different data sources and different methods were compared.The RF algorithm integrating UAV hyperspectral images and WV-2 multispectral images performed the best with an overall classification accuracy of 95.89%,which was 0.54% higher than the SVM(RBF)algorithm,with little improvement in accuracy compared to that using UAV hyperspectral images.Regarding the classification methods,the RF algorithm outperformed the SVM algorithm in terms of accuracy,stability,and adaptability.As a vital ecosystem,species classification studies of mangroves are beneficial for wetland management and sustainable development.This study confirmed the feasibility and development potential of mangrove species classification by using high-quality data sources and effective methods.Future research will focus more on the application of high accuracy vegetation species classification results and work towards improving sustainable mangrove management.