关键词: 林分类型, 可见光遥感, 无人机, 特征提取, 特征优选, 分类

Abstract:

Objective: The classification of bamboo forest,coniferous forest and broad-leaved forest using visible data acquired by unmanned aerial vehicle(UAV) remote sensing could expand the application scope of UAV visible light remote sensing data in forest resource survey. Method: The aerial image acquired by UAV only containing the red,green and blue three-band information was preprocessed to generate digital orthophoto map(DOM)and digital surface model(DSM)with a spatial resolution of 0.1 m. Six categories of features were extracted from DSM and DOM,including height features,spectral features,common vegetation indices,HSV color components,texture features extracted based on HSV color component,and extended morphological multi-attribute profiles(EMAPs). Random forest_recursive feature elimination(RF_RFE)was used to determine the optimal feature set. Eight groups of experiments were constructed according to different types of features and selected feature subsets,and then the forest stand were classified by random forest classifier(RFC). The evaluation classification results of confusion matrix were established using real ground images obtained by interpretation. Result: 1) Forest stand classification using spectral characteristics alone was not satisfactory,the overall accuracy was 65.68%,and the Kappa coefficient was 0.53. Based on the spectral features,other features could be introduced to improve the overall accuracy except vegetation indices. 2) Among the 11 features optimized by the RF_RFE,there were 5 EMAPs features,3 HSV texture features,1 height feature,1 vegetation index and 1 HSV color component. These 11 features obtained the highest classification accuracy in 8 groups of experiments,with an overall accuracy of 81.05% and a Kappa coefficient of 0.73. 3) A good classification result could be obtained by applying the multi-feature preferred method to the visible light drone images of different resolutions. When the resolution was 0.3 m,the classification accuracy was the highest,the overall accuracy was 82.46%,and the Kappa coefficient was 0.75. Conclusion: The research showed that the RF_RFE algorithm could synthesize the most favorable features of multi-type features in forest stand type classification,and thus improve the classification accuracy. This study could provide a reference for the extraction of forest stand information in forest resource inventory.

Key words: forest stand types, visible light remote sensing, unmanned aerial vehicle(UAV), feature extraction, feature selection, classification