Research On Canopy Nitrogen Content Retrieval Of Mangrove Species Based On Satellite-UAV-Ground Multi-Source Remote Sensing

Mangroves,which grow in the tropical and subtropical sea-land ecotone,are one of the most productive coastal wetland ecosystems.They play an important role in water purification,carbon sequestration,emission reduction,biodiversity protection,and ecosystem balance.Nitrogen is an indispensable nutrient element for the growth and development of mangrove communities,while it is also one of the most crucial elements leading to mangrove wetland pollution and water eutrophication.Canopy nitrogen content(CNC)of mangroves is not only an important part of wetland nitrogen cycle,but also significantly affects the photosynthesis of different mangrove species.The CNC retrieval of mangroves not only has important application value for diagnosing the physiological status and growth trend of mangroves,as well as evaluating the health level of wetland habitat,but also has vital theoretical significance for protecting plant diversity.In addition,among the monitoring technologies of CNC of mangroves,multi-source remote sensing technology could significantly improve the research accuracy and work efficiency,as it is not limited by spatial scale.It provides technical support and scientific basis for fine ecosystem management.In the study,the Zhangjiang Estuary Mangrove National Nature Reserve was chosen as the research area.Integrating the unmanned aerial vehicle(UAV)multi-spectral images,Sentinel-1 radar satellite images and ground survey data,the object-oriented optimal segmentation scale was determined using the local variance change rate of different segmentation scales.On this basis,the features of multi-source remote sensing images were mined deeply,and the optimal feature combination for mangrove species identification was determined by the distance separability criteria.Then,the U-net deep learning algorithm and its accuracy evaluation were carried out on the image identification of mangrove species.Furthermore,four typical machine learning methods,i.e.K-nearest neighbor(KNN),CART decision tree,Bayes,and Random forest(RF),were selected for comparison to evaluate the advantages and disadvantages of the image identification method.The Genetic algorithm-BP neural network method was used to construct the CNC retrieval model of mangrove species,thereby realizing the spatial distribution mapping.Finally,the CNC spatial heterogeneity analysis of different mangrove species was carried out.The main conclusions are as follows:(1)Based on object-oriented multiresolution segmentation,the change rate of local variance and segmentation effect of different segmentation scales were analyzed to select the segmentation scale,shape heterogeneity weight,and compactness weight as 1800,0.1,and 0.5,respectively.It indicated that the segmentation effect is highly consistent with the boundaries of mangrove species.The optimal image feature combination was selected quantitatively based on the distance separability criterion,including 25 features,with 12 spectral features,6shape features,5 texture features,and 2 height features.The optimal image features could effectively represent mangrove species with the highest intra-species similarity and the lowest inter-class similarity.(2)Based on the Sentinel-1 radar image and the vegetation index derived from UAV multi-spectral data,a hierarchical strategy was adopted to identify mangroves and non-mangrove areas in the study area.Combing with the object-oriented method,U-net deep learning method was used to identify mangrove species information.Then,the identification results of U-net deep learning method were compared with those of KNN,CART,Bayes and RF methods.The accuracy evaluation showed that the U-net method had the highest identification accuracy with 92.02%.The overall accuracy increased by 12.11%to 22.79%compared with other identification methods,and Kappa coefficient increased by 0.20 to 0.36.The results showed that the object-oriented U-net deep learning method could effectively identify different mangrove species,which had obvious advantages in identification accuracy compared with other typical machine learning methods.(3)The CNC sensitivity features of mangroves were selected based on genetic algorithm,and then 11 selected sensitivity features were introduced into BP neural network constructing the CNC retrieval model of mangroves.The results showed that the estimation accuracy of BP neural network for Avicennia marina(AM),Kandelia candel(KC),and Aegiceras corniculatum(AC)are 93.8%,87.83%and 95.55%,respectively.The CNC of AM,KC,and AC range from 3.59 to 8.94 g/m~2,1.49 g/m~2to 8.09 g/m~2and 2.97 g/m~2to 4.46 g/m~2,with the mean values of 6.04 g/m~2,3.76 g/m~2and 3.67 g/m~2,respectively.The CNC spatial distribution of mangroves increased from northwest to southeast in the study area,and the retrieval mean value is 4.00 g/m~2.In addition,CNC of mangroves had synthetically affected by species type,elevation and offshore distance.The study could provide technical support for image identification and ecological parameters retrieval of mangrove species based on multi-source remote sensing images.It could also provide theoretical references for ecological restoration and fine management of mangrove wetlands.