Study And Evaluation Of Desert Grassland Fractional Vegetation Coverage Based On UAV Hyperspectral Remote Sensing

Monitoring and research of fractional vegetation coverage is an important part of the evaluation of grassland degradation.It provides a sound basis for understanding the evolution of the ecosystem and formulating grassland management strategies.Affected by climate change,overgrazing,weed invasion and other factors,we still face great pressure of desert grassland degradation and significantly declined vegetation productivity.Grassland ecosystems in some areas are extremely fragile in ecosystem sustainability and resilience to interference,showing a trend towards bare land and desertification.According to Grading Index for Degradation,Desertification and Salinization of Natural Grasslands(GB 19377-2003),Grassland desertification takes the form of reduced vegetation community coverage,decreased vegetation types,and increased area of bare land,etc.Therefore,the monitoring and inversion of desert steppe ecosystem coverage can provide a theoretical basis for the evaluation of grassland degradation,giving it great practical significance for the study of d esert grassland degradation and ecological environment protection.Real-time,rapid and accurate collection of grassland coverage information is the basis for grassland monitoring and research.The traditional small-scale collection of ground vegetation coverage is greatly affected by weather,time,and topography.Moreover,it is costly,time-consuming and labor-intensive,and could only collect information about the vegetation change,composition and distribution in a specific area.Artificial satellites,on the other hand,can be used on a large scale to collect images covering vast areas of the earth’s surface,but the collection is slow,and the images are relatively sketchy.UAV + hyperspectral remote sensing images boast short cycles,wide coverage,rich data sources,rapid and repeated deployment,saving of time and labor and other advantages.The combination produces high spectral resolution,high spatial resolution and multi-temporal data in mesoscale with both accuracy and efficiency up to the standards.Therefore,it is an unparalleled data source for studying fine-scale surface coverage.It plays an important role in monitoring grassland degradation in relatively small areas and is becoming an excellent complement to traditional ground monitoring and aerial and satellite remote sensing.We introduce the UAV + hyperspectral remote sensing mode to grassland coverage monitoring,so as to figure out methods for estimation of the distribution and coverage of desert grassland vegetation based on the spectral feature data of ground objects and thereby provide data reference and basis for the decision-making and treatment of grassland degradation.The distribution of plant communities is sparse,low and staggered in coverage.Addressing such existing problems as pixel aliasing and low coverage estimation accuracy for ground objects,this research did the following work with UAV +hyperspectral remote sensing images as the data source and Gegentala in Siziwang Banner of Inner Mongolia as the research area: accurate calculation of desert grassland coverage,in-depth discussion of the capability and applicability of regression model,mixed pixel decomposition and deep learning in calculating coverage,study of the inversion method of desert grassland coverage in an all-round and systematic manner based on UAV hyperspectral data,and comparison and validation of the results of different methods.The main research content and conclusions are as follows:(1)The green light continun removal soil adjustment vegetation index(G＿CR＿SAVI)and the green light continun removal normalized difference vegetation index(G＿CR＿NDVI)were proposed.Based on a comparison of the applicability of 4traditional vegetation indexes—normalized difference vegetation index(NDVI),ratio vegetation index(RVI),difference vegetation index(DVI),and soil adjustment vegetation index(SAVI)—in calculating the vegetation coverage of desert grassland,the most sensitive characteristic bands were screened out through spectral enhancement continuum removal,simple-band autocorrelation selection,and MATLAB programming.The vegetation index calculated by the optimal band combination was substituted into the dimidiate pixel model to invert the vegetation coverage.Moreover,20 sets of data were selected for precision verification.The estimated precision was in order: G＿CR＿SAVI>G＿CR＿NDVI> SAVI> DVI> NDVI> RVI.The estimation accuracy of the improved vegetation index is about 4% higher than that of the traditional vegetation index.The applicability and sensitivity of the regression model statistics(vegetation index)were also detected in estimating coverage based on UAV hyperspectral data in desert grassland.(2)An improved fully constrained least squares algorithm(IM-FCLS)based on the linear unmixing model was proposed.In view of the common phenomenon of pixel spatial aliasing in spectra of desert grassland vegetation,FCLS based on the linear mixing model was introduced for estimation of photosynthetic vegetation coverage and non-photosynthetic vegetation coverage.It was noticed through MNF rotation,PPI endmember selection,N-dimensional visualization and FCLS abundance estimation that,the FCLS algorithm could easily cause multiple iterations,large time consumption or even iterative infinite loops when there was a non-negative minimum in the non-negative iteration process.After setting the endmember abundance corresponding to the negative value with the largest absolute value to 0 by using the IM-FCLS algorithm,and removing this endmember spectrum,the coverage estimation accuracy was slightly improved compared with the FCLS algorithm,with less time consumed.(3)A deep learning algorithm for desert grassland coverage was established based on UAV hyperspectral remote sensing.First of all,datasets of vegetation and non-vegetation(bare soil,shadow,etc.)were prepared and sorted out by four deep learning network models,namely VGG16,Res Net18,3D-VGG16,and 3D-Res Net18.VGG16 and Res Net18,two classical models,showed better performance in classifying vegetation and soil in remote sensing images of desertification.3D-VGG16 and3D-Res Net18 models improved by the 3D convolution kernel also reached good potential for classification of vegetation,soil and small samples in remote sensing images of desertification,with the overall classification accuracy 1-2% higher than that of Vgg16 and Res Net18.According to optimization and comparison of parameters such as the number and size of convolution kernels,Batch size,activation function,and proportion of training samples,the model with the best classification performance for the coverage datasets is 3D-Res Net,with estimation accuracy up to 97.74% and the RMSE of coverage estimation at 0.016.It achieves intelligent,high precision,fast and accurate estimation of desert grassland coverage.(4)Evaluation of grassland degradation with vegetation coverage calculated by the3D-Res Net model as the index was achieved.Vegetation coverage in 89 sets of hyperspectral images of areas with different grazing intensities was calculated in the spatial dimension.The determined grassland degradation grades were basically consistent with the actual characteristics of the surface degradation areas.Among 89 verification points,81 were correctly estimated in the degradation grade,the accuracy being 91.01%.In addition,vegetation coverage of 675 sets of hyperspectral images during the growing season in 2017～2021 was calculated in the temporal dimension,which found a certain downward trend of vegetation coverage over time in the research area.This verified the practicability and accuracy of 3D-Res Net in coverage calculation for the evaluation of grassland degradation.This thesis obtains hyperspectral images of Gegentala in Inner Mongolia in 5 years through the UAV hyperspectral remote sensing platform.It enriches the spatial and temporal scale of research on remote sensing of grassland desertification and explores necessary hardware integration for remote sensing of grassland desertification.Reliable methods proposed herein for calculation of desert grassland coverage may serve as a theoretical basis and technical support for graded evaluation and treatment of degradation taking grassland vegetation coverage as a necessary monitoring indicator.