Nitrogen Content Retrieval Of Wetland Plant Community Canopy Based On Unmanned Aerial Vehicle Hyperspectral Remote Sensing

Wetland plants are an important indicator of wetland degradation and restoration.Monitoring the nitrogen content in the canopy of wetland plant communities is an important indicator for evaluating the growth status of wetland plants.Monitoring wetland plant community distribution and canopy nitrogen content based on hyperspectral remote sensing technology is of great significance for timely control of the wetland plant community distribution pattern and implementation of precise management.It can provide real-time,efficient and non-destructive data acquisition technology for mapping the spatial distribution of wetland plant communities and retrieving canopy nitrogen,with the rapid development and application of UAV hyperspectral remote sensing technology.In this study,taking the dominant plant community in the core area of the Momoge wetland as the research object.The UAV-borne imaging hyperspectral system was used to obtain the canopy hyperspectral image of the wetland plant communities.Use object-oriented image analysis(OBIA)and machine learning[Random Forest(RF),Convolutional Neural Network(CNN)and Support Vector Machine(SVM)]classification method to complete the spatial distribution map of wetland plant communities.Based on these,it used that correlation analysis to select the characteristic bands of canopy nitrogen content,combined with three modeling methods of Partial Least-squares Regression(PLSR),Random Forest(RFR),Back Propagation Neural Network(BPR)to constructed hyperspectral estimation models for canopy nitrogen content of wetland plant communities.The results show that:(1)By designing and carrying out field experiments,we obtained UAV hyperspectral image data,ground survey data and measured data of wetland plant community canopy.After the image preprocessing process,the hyperspectral image data covering the research area is obtained.These experiments provide a data basis for the subsequent selection and extraction of canopy spectrum features,feature sensitivity analysis,and classification algorithm research of wetland plant communities.(2)Study the spectral feature extraction and preferred methods.In order to solve the problem of feature redundancy and dimensionality disasters in hyperspectral data,this paper studies the hyperspectral characteristics of typical wetland plant communities in the study area from multiple angles.Firstly,the preprocessed hyperspectral data were processed by derivative operation,logarithmic operation,logarithmic derivative operation and continuum removal transformation to highlight the differences of reflection spectrum and absorption spectrum characteristics of each community type.Then,the feature extraction method of principal component analysis was used to perform spectral dimensionality reduction on various transformations.A total of 71 characteristic bands were selected from various spectral characteristics.(3)A classification method was proposed by fusing OBIA and machine learing method for wetland plant community.Based on the dimensionality reduction data of hyperspectral images,object-oriented and machine learning classification algorithms(RF,CNN and SVM)were used to extract wetland plant communities and compared classification accuracy.The overall classification accuracy of random forest,convolutional neural network and support vector machine classification results at object scale are 87.75%,83.31%and 80.29%,respectively.And the Kappa coefficients are 0.86,0.83 and 0.81,respectively.Compared with the classification accuracy of the same method at pixel scale,it has increased by 15.06%,8.98%and10.87%.The results of classification comparison experiments showed that OBIA-RF was the best scheme for wetland plant community classification.The study area was dominated by phragmites australis community and typha latifolia community.According to the best classification scheme,the area of phragmites australis community is 295 hectares,accounting for 57%of the total area of the study area.It was widely distributed throughout the study area.The area of the typha latifolia community was 84 hectares,accounting for 16%of the total area.It was mainly distributed in the southern and northern regions of the study area.The area of suaeda glauca community was 79 hectares,accounting for 15%of the total area.It was mainly distributed in the west and north of the study area.The area of scirpus triqueter community was 24 hectares,accounting for only 5%of the total area.And it was scattered among the phragmites australis community and typha latifolia community.(4)Based on the hyperspectral data,three spectral variables of the wetland community canopy were obtained:sensitive spectral band information,trilateral spectral characteristic variables and hyperspectral vegetation indices.And sensitive bands or spectral characteristic variables were selected through correlation analysis.Finally,the canopy nitrogen content of four wetland plant communities was estimated by using partial least square regression(PLSR),random forest regression(RFR)and BP neural network regression algorithm combined with sensitive spectral variables.The results shown that the BP neural network model based on spectral feature variables and hyperspectral vegetation index could provide more accurate results in estimating the canopy nitrogen content of four wetland plant communities.The BP neural network inversion model based on spectral feature parameter combinations at714 nm(phragmites australis)and 710 nm(typha latifolia)was the best estimation model for the nitrogen content in the canopy of the phragmites australis community and the typha latifolia community.The R² of regression equation with training dataset,root mean square error(RMSE)values were 0.87,0.89,0.29 and 0.23.The R² of regression equation with validation dataset,root mean square error(RMSE)values were 0.83,0.85,0.19 and 0.27.The BP neural network inversion model based on spectral feature parameter combinations(red edge amplitude(Dr)at 706 nm,red edge position(PDr),yellow edge area(SDy)at 618 nm,red edge area(SDr),red valley area(SR₀)was the best estimation model for the nitrogen content in the canopy of the suaeda glauca community.The R² of regression equation with training dataset,root mean square error(RMSE)values were 0.81,0.81,0.46 and 0.25.The BP neural network inversion model based on vegetation index combinations was the best estimation model for the nitrogen content in the canopy of the scirpus triqueter community.The R² of regression equation with training dataset,root mean square error(RMSE)values were 0.76,0.75,0.36 and 0.35.On this basis,the UAV hyperspectral image and the optimal inversion model were used to estimate the nitrogen content of the four wetland plant communities'canopy and make the spatial distribution map of the nitrogen content.For phragmites australis community and typha latifolia community,the spatial distribution pattern of canopy nitrogen content presented a trend of high in the surrounding regions and low in core regions.The canopy nitrogen content of suaeda glauca community was higher in the north and south regions and lower in the core regions.The canopy nitrogen content of the phragmites australis community ranges from 7.90 to 48.47 g/kg,with an average value of 12.96 g/kg;the canopy nitrogen content of the typha latifolia community ranges from 6.67 to 36.37 g/kg,with an average value of 15.39 g/kg;The canopy nitrogen content of suaeda glauca community ranges from 7.03 to 42.18 g/kg,with an average of 17.393 g/kg;the canopy nitrogen content of scirpus triqueter community ranges from 6.55 to 23.58,with an average of 15.98 g/kg.This research has realized the non-destructive,fast and efficient remote sensing monitoring of nitrogen content from points to regions.This provides a scientific basis for monitoring the physical and chemical parameters of wetland plant communities and implementing effective management.