Research On Extracting Vegetation Information In Karst Wetland Based On UAV Images And Deep Learning

Wetland vegetation is one of the three elements of wetland and plays an important role in maintaining the wetland process.This paper took the core area of Guilin Huixian Karst Wetland Park as the research area.This paper made UAV DOM image and textural feature,DSM as data sources.The object-oriented random forest classification model combined with visual interpretation was used to produce karst wetland label data.used the core area of Guilin Huixian Karst Wetland Park as the research area,used drone DOM images and textural features,and DSM as data sources,and uses an object-oriented random forest classification model combined with artificial visual interpretation to produce karst wetland label data.This paper used Seg Net,PSPNet,RAUNet and Deep Lab V3 plus networks as methods to construct a deep learning model for extracting vegetation information in karst wetlands.In order to improve the accuracy of the model to identify vegetation information,and further use soft-voting method,hard-voting method,and optimal fusion method to construct deep learning fusion model about remote sensing.The experiment mainly evaluated the accuracy of deep learning in the extraction of karst wetland vegetation information.And it explored the influence of DSM and textural features on the four deep learning models to extract vegetation information in karst wetland,and evaluated the best way of the three fusion methods.The research conclusions are as follows:(1)In the Seg Net,PSPNet,RAUNet and Deep Lab V3 plus multi-classification models constructed by the DOM dataset,the pixel-based and sampling-based of F1-scores and Kappa coefficient are all higher than 0.70.This showed that the deep learning model can better identify vegetation information in karst wetlands.Among them,the Deep Lab V3 plus model and RAUNet model had the highest accuracy in identifying karst wetland vegetation.The result of Deep Lab V3 plus model had the pixel-based of F1-scores and overall accuracy reaches about0.78.The result of the RAUNet model had the sampling-based of F1 score and the overall accuracy is about 0.88.(2)In the single classification model of DSM and textural feature data sets,textural feature could improve the accuracy of four deep learning models to identify water hyacinth,lotus and cultivated land.At the same time,it affected the accuracy of the model's identification of karst herb,linden tree-vesicle-bamboo information.DSM images could improve the extraction of karst wetland vegetation information from the four deep learning models.Among them,DSM data improved the lotus identification information best,the pixel-based and sampling-based of F1 score increases by about 0.07.Textural features improved the information of cultivated land and water hyacinth the best,and the pixel-based and sampling-based of F1 score was improved by about 0.05.(3)The three fusion deep learning models were better for extracting karst wetland vegetation information.The fusion model based on the optimal-fusion method extracted the geometric and attribute information of karst wetland vegetation was best.The pixel-based F1 score,Kappa coefficient,and overall accuracy were 0.8702.0.8304,0.8684.And the samplingpoints-based F1 score,Kappa coefficient,overall accuracy were 0.9445,0.9310,0.9444.(4)The overall accuracy and Kappa coefficient of the RAUNet multi-classification model results are 0.0242 and 0.0308 lower than the object-oriented random forest classification model.The fusion model based on the optimal fusion method has higher recognition accuracy than the object-oriented random forest classification model,and the overall accuracy and Kappa coefficient were higher than 0.0424 and 0.0510 respectively.It showed that the object-oriented random forest model is better than the deep learning model.But the optimal fusion method is a way to effectively improve the deep learning model to identify vegetation information in karst wetlands.