Grassland Biomass Inversion And Functional Group Identification Based On Hyperspectral Remote Sensing

The degradation of alpine grassland has become an important ecological problem for human beings.The alpine grasslands in China are mainly distributed in the Qinghai-Tibet Plateau region,which is an important base for livestock production and an important ecological protection barrier in China.In recent years,due to the influence of human activities and extreme climate,alpine grasslands have undergone extensive degradation,both in terms of reduced forage production capacity and in terms of changes in grassland community structure.The biomass and vegetation functional groups are two important indicators reflecting the basic characteristics of vegetation,which are the main way to understand the material cycle and energy flow of grassland ecosystem,and the changes of both affect the stability of grassland ecosystem and bind the socio-economic development.Hyperspectral remote sensing data contains rich waveband information,which can make fine observation of grassland vegetation and realize biomass inversion and functional group identification.However,hyperspectral data are rich in noise and redundant in information,which can easily lead to overfitting of inversion models.Effective spectral feature extraction and selection mechanisms need to be designed to minimize the correlation between bands.Existing spectral downscaling methods mainly rely on manual feature extraction and then achieve spectral screening based on the contribution of features.However,such methods are time-consuming and laborious,lacking in mechanics and poor in universality,and cannot effectively fuse the rich spectral and spatial information in hyperspectral images.Taking the alpine grassland in Henan County,Qinghai Province as the research object,we use the feature spectrometer and unmanned aerial hyperspectral equipment for vegetation observation,and realize the inversion of grassland biomass and functional group identification based on machine learning algorithm.For the inversion study of above-ground biomass,four feature selection techniques(m RMR,T-test,Relief F,Lasso)are used to evaluate the importance of each variable and filter out sensitive bands,and machine learning regression algorithms(SVM,GPR,LSB)are used for above-ground biomass inversion,and then the specific role of features in model prediction is analyzed based on interpretability theory.For the identification of functional groups of grassland plants based on multi-source hyperspectral data,a data fusion classification framework based on Multi-branch Convolutional Neural Network(MBCNN)is proposed in this paper.The method is divided into three steps:firstly,a one-dimensional convolutional neural network(1D-CNN)is used to extract the feature spectra of grassland features;then,a three-dimensional convolutional neural network(3D-CNN)module is constructed to extract features from the hyperspectral images of UAVs in two wavelength ranges(Visible and Near-infrared)respectively;finally,the extracted three sets of hyperspectral features are stitched together and combined according to the functional group categories(Gramineae,Salicaceae,Leguminosae,Asteraceae,Rosaceae)for group identification.The results of the study were as follows.(1)The above-ground biomass of natural grasslands showed an approximately normal distribution,and the correlation between biomass and different spectral bands varied greatly.The above-ground biomass of grasslands increased with seasonal succession,but the biomass of grazed areas increased at a lower rate than that of ungrazed areas;the overall biomass ranged from 9.35～220.54 g,with an average biomass of about 80.58 g.The visible(350～500 nm and 600～700 nm)and near-infrared bands(700～1350 nm)had the highest correlation with biomass and were important variables for remote sensing analysis of grasslands.(2)Based on feature selection and machine learning for biomass inversion,Lasso and m RMR feature selection algorithms can effectively reduce the data dimensionality,and the inclusion of environmental factors in the dataset can improve the accuracy of the grassland AGB inversion model.Lasso and m RMR can drastically reduce the number of features in the dataset,which effectively reduces the multicollinearity among variables and makes the modeling performance improved.The addition of two environmental factors,growing season and grazing condition,greatly improved the R~2of the grassland AGB estimation model(0.17～0.39)compared to the original spectral features.Compared with the modeling results of other regression models(SVM,LSB),the inverse model of grassland AGB constructed using the GPR algorithm was more accurate.The best inverse model of AGB was the GPR model built based on the multi-source data feature variables screened by m RMR with an accuracy of R~2=0.54 and RMSE=19.571 g.(3)The five functional groups(Gramineae,Salicaceae,Leguminosae,Asteraceae,and Rosaceae)of the grassland were widely distributed in the alpine grassland,with Gramineae being the dominant functional group(accounting for 73%).The linear correlations among functional groups were weak,with correlation coefficients ranging from-0.4 to 0.35,mainly positive.The spectra of grassland functional groups have the typical characteristics of vegetation spectra,with obvious peaks and valleys,and the differences in reflectance between different types of spectral data in the near-infrared band are more obvious,but the overall differences are small.(4)In the functional group recognition model constructed based on multi-source data and deep learning,the accuracy of modeling using multi-source hyperspectral data is higher than that of feature spectral data.The multi-branch convolutional neural network can fully exploit the rich spectral and texture information in the grass canopy hyperspectral data and fuse the observation data at both near-ground and unmanned aircraft scales for effective analysis,and the accuracy reaches 81%,53%,75%,62%,and 62%in the recognition tasks of five functional groups,respectively,and the improvement of accuracy is in the range of 5%～10%compared with the results of geo-spectral modeling.Taken together,the use of convolutional neural network models to analyze the spectral data of plant functional groups has achieved good results in the identification of grass functional groups by improving the feature extraction ability while also reducing the number of model parameters.