| Forest ecosystems cover approximately one-third of the Earth’s land surface and are an important component of terrestrial ecosystems,as well as the largest carbon sink on land.The distribution of tree species is an important factor in forest ecosystem research.Obtaining accurate information on tree species distribution is not only significant for sustainable forestry management but also contributes to a better understanding of the ecological status of forests.Traditional vegetation surveys are often time-consuming,laborious,costly,and pose some safety hazards.Although highresolution satellite data can provide remote sensing data at the meter or even sub-meter level,it faces problems such as poor timeliness,high cost,and the impact of cloud and fog weather,making it difficult to meet the requirements of dynamic observation.Unmanned aerial vehicle(UAV)platforms are not affected by cloud cover,have the advantages of maneuverability,flexibility,and low flight costs.They can not only obtain high spatial resolution remote sensing data but also adapt to complex geographical environments,making them widely used in the field of forestry remote sensing.However,most tree species classification studies are often concentrated in artificial forests or plantations,with little consideration of more widely distributed natural forests or secondary forests.Meanwhile,the spatial scale and temporal characteristics of UAV images in tree species classification are not clear,and many researchers tend to rely on empirical acquisition of UAV images and are unable to make accurate vegetation classification in large-area forests.Therefore,to solve the above problems,this thesis takes subalpine coniferous natural forest and warm temperate deciduous broad-leaved secondary forest in western Sichuan as the research object,and obtains high-resolution UAV images and monthly-scale observation images based on the whole growing season,and combines convolutional neural network to study the spatial scale and temporal characteristics.Additionally,the high resolution,stability,and flexibility of multi-rotor UAVs and the large-scale and high-efficiency characteristics of fixed-wing UAVs were combined to identify tree species in large forest areas.(1)This study combines ultra-high resolution UAV images and convolutional neural networks(CNNs)for tree species classification,finds a better cropping size model and constructs vegetation classification models based on UAV images of different resolutions for suitability spatial scale analysis.The results showed that combining visible UAV imagery with CNN models can achieve high accuracy in tree species classification.The optimal cropping size was found to be 256×256 pixels at a spatial resolution of 5cm,with an overall accuracy of 93.21% and a Kappa coefficient of 0.90.The increase of spatial resolution can improved classification accuracy,but the ability of this improvement decreased greatly when the spatial resolution reached0.1m.When spatial resolution increased from 15 cm to 10 cm,the overall accuracy of the model increased by 14%,and the Kappa coefficient increased by 0.15.However,when spatial resolution increased from 10 cm to 5cm,the overall accuracy of the model only increased by 2%,and the Kappa coefficient increased by 0.03.Vegetation types that are not representative in the region are more affected by spatial resolution than dominant tree species.This indicated that the spatial resolution of images should be selected based on the characteristics and number of categories of classified features,and blindly pursuing high spatial resolution may not necessarily bring significant improvements in model classification accuracy.(2)By continuously monitoring the dominant tree species at a monthly time interval throughout the entire growing season,using both visible and multispectral drone imagery,we clarified the differences in capabilities between the two and revealed the changes in the classification accuracy of UAV imagery at different times.The results showed that the classification accuracy of visible imagery varied widely in different months(highest Intersection over Union,Io U,of 83.72% in April and lowest Io U of 76.81% in June),while the accuracy of multispectral imagery was consistently better in all months(Io U was always greater than 81%,and the highest accuracy occurred in June).The model accuracy of the multispectral imagery was higher than that of the visible imagery under most months,especially in June when it was nearly6% higher than that of the visible image in Io U.However,in April and October,the model accuracy of the visible imagery is close to or even slightly higher than that of the multispectral imagery.Vegetation indices were also added to the model to increase information,and the results only showed improvement in the model accuracy of visible imagery during the summer,but there was still a certain gap compared to the model accuracy of multispectral imagery during the same period.This indicated that the observation time has a certain influence on the accuracy of tree species identification,especially for visible images;and as the interspecies differences caused by month changes in the study area gradually increase,the advantage of multispectral images over visible images gradually weakens or disappears.(3)We proposed a method for identifying a large range of dominant tree species under forests with complex topography and high heterogeneity by exploiting the synergy between fixed-wing UAVs and multi-rotor UAVs.Multi-rotor UAVs were used to capture high-resolution images of small areas during different seasons,and combined with CNNs to obtain a highly accurate classification model.To improve identification of tree species from fixed-wing UAV imagery,the predictive capability of the multi-rotor classification model was used and combined with visual interpretation data to expand sample annotation,and finally with convolutional neural networks for tree species identification.The results show that this method is able to perform tree species recognition in complex forest environments over a large area of forest(F1 score of 92.93% for the fixed-wing model).This method solves the problem of low accuracy of fixed-wing UAV images in large-range tree species identification,and combined the stable data quality of multi-rotor UAVs with the data collection efficiency of fixed-wing UAVs to achieve high-precision tree species identification over large areas,which has significant implications for regional and large-scale forest distribution surveys.Overall,this thesis used different types of UAVs to continuously monitor typical areas of subalpine coniferous natural forests and warm temperate deciduous broadleaved secondary forests,combined with CNNs for tree species identification research to explore the spatial and temporal characteristics of UAV remote sensing,and achieved large-area tree species identification with the synergy of multi-rotor and fixed-wing.The conclusions of this thesis can not only provide some guidance for the spatio-temporal observation of UAVs in forests,but also provide theoretical support and application reference for UAV synergy. |
PDF Full Text Request