Feature-level Fusion Between Hyperspectral And Synthetic Aperture Radar Data For Tea Plantation Mapping

Due to the high economic value of tea,the area of tea plantations in China has been increasing in recent years.Therefore,the accurate mapping of tea plantations is significant for government decision-making and environmental protection.There are some problems in the previous researches about tea plantation mapping using remote sensing technology,such as low accuracy,atypical features,and data source challenging to obtain.Hyperspectral and Synthetic Aperture Radar(SAR)data have the advantages of easy acquiring and high performance,and they have been widely used in extracting surface information.However,there are few reports on their application in tea plantation mapping.Extracting valuable features from hyperspectral data and effectively combining two data for tea plantation mapping still needs further study.To address these problems,this study selects the Wuyishan area,Fujian Province,as the study area to research tea plantation mapping based on hyperspectral and SAR data.The main research results as follows:(1)This paper studies the feature extraction of tea plantations based on hyperspectral and SAR data.The extremum bands,principal component bands,and texture of the hyperspectral data are firstly extracted.Then,three vegetation indexes about slope and area are established based on the hyperspectral spectral characteristic of tea plantations,effectively improving the separability of tea plantations.Finally,the backscattering information and the corresponding texture,and the polarization decomposition information are calculated from the full-polarimetric SAR data.(2)Based on the locality preservation projection and Subspace Fusion(Sub Fus),this paper proposes a new feature-level image fusion method(LPPSub Fus).Then,the support vector machine(SVM)and neural network algorithm are used to compare the performance of LPPSub Fus,Sub Fus,and three traditional pixel-level image fusion methods in tea plantation mapping.The results show that the LPPSub Fus tends to be more advantageous in classification accuracy and visual effect.The LPPSub Fus using SVM algorithms achieves the highest accuracy of 95.21%,with the producer and user accuracy of tea plantations and forests more than 90%,demonstrating the effectiveness of LPPSub Fus.(3)It is also found that the classification using only hyperspectral data or SAR data has many shadows and speckle.The classification obtains effective improvement after feature-level image fusion,showing the potential of the feature-level image fusion of hyperspectral and SAR data in tea plantation mapping.The research also suggests that the feature-level image fusion can reduce the noise in the classification and improve the classification accuracy of tea plantation mapping.Moreover,it has the advantage of data dimensionality reduction and has excellent performances in the two classification algorithms.Therefore,the feature-level image fusion between hyperspectral and SAR data can meet the requirement of accurate tea plantation mapping.