Multi-Scale Remote Sensing Inversion Of Biomass Component Of Larix Olgensis

As an important indicator to measure forest ecosystem productivity,forest biomass plays an important role in explaining the mechanism of global carbon cycle,coping with global climate change and studying ecosystems.The estimation of forest biomass is traditionally based on high-density sample plot survey and realized by constructing allometric growth model.It has the advantages of simple principle,strong applicability and high estimation accuracy.However,it consumes a lot of manpower and material resources,and leads great disturbance to forest ecosystems,thusit is difficult to achieve long-term and macro dynamic monitoring of forest biomass.Remote sensing technology has many successful cases in the estimation of forest biomassand shows irreplaceable advantages,because of its super continuous detection and extraction capability in dealing with various complex surface,large area and long time span..However,there are some limitations in the practical application of single remote sensing means,such as the low accuracy of biomass estimation caused by the saturation of optical remote sensing data;the small number of microwave radar data sources and the vulnerability to terrain fluctuations;the discrete attributes of lidar and the high cost of airborne lidar,which make it difficult to achieve large-scale continuous monitoring.Therefore,it is of great significance to study how to overcome shortcomings of single remote sensing means,comprehensively use the integrated air-space-ground monitoring method,solve key technologies such as integration,rapid and comprehensive processing of multi-source remote sensing information,and achieve large-scale,rapid and accurate access to forest biomass.This study takes the L.olgensis plantation in Heilongjiang Province as the research object.A complete set of plan in forest biomass estimation was put forward by by introducing a deep learning method based on synthetically utilizing data in ground survey,aerial remote sensingand satellite remote sensing.This achieves a wide range,rapid and accurate acquisition of forest biomass from different scales such as single-wood-forest-operating unit-regional macro.The main conclusions are as follows:1?Based on the data of 40 sample plots and 64 sample trees of L.olgensis plantation,in the case of considering and not considering the age of the forest,the dummy variable and the non-linear likelihood-independent regression method were combined to construct the one-dimensional compatible biomass models of individual tree and stand levels respectively.The R_adj²of each biomass model was basically greater than 0.91 at single tree level,and the average prediction accuracy was more than 92%.At stand level,the R_adj²of each biomass model is more than 0.98,and the average prediction accuracy is more than 97%at stand level.The addition of age factor can significantly improve the fitting effect,prediction accuracy and model stability of stand biomass model,but has little effect on individual tree biomass model.2?A remote sensing inversion model of L.olgensis biomass was constructed by using multivariate linear regression and Stochastic Forest methods,using the ground sample data and airborne Li DAR point clouds acquired simultaneously as data sources.Independent variables extracted by airborne Li DAR were significantly correlated with biomass,which generally showed significant?P<0.05?or extremely significant?P<0.01?.Multivariate linear regression and random forest can be used to construct remote sensing inversion model of Larch biomass.R²of the model is higher than 0.91,and both have smaller r RMSE and TRE.The advantage of multiple linear regression is that the model is simple and easy to use,and the independent variable of Li DAR can explain biomass more clearly.Random forest model has better fitting effect and generalization ability.The improved-RF method is proposed to improve the stochastic forest model for the first time.The results show that the improved-RF method can effectively avoid the common over-fitting problem of stochastic forest,and greatly reduce the number of independent variables without loss of prediction accuracy.3?Based on the biomass inversed by airborne Li DAR and the spectral features,vegetation index and texture features extracted from GF-1 remote sensing data.Random forest and cyclic neural network methods were used to construct remote sensing extrapolation models of biomass component of L.olgensis plantation.Cyclic neural network and stochastic forest model can explain the complex non-linear relationship between remote sensing factors and biomass.R²of main biomass models such as whole plant,aboveground and tree roots are better than 0.7.Cyclic neural network shows stronger ability in preventing biomass underestimation and reducing observed data structure.The cyclic neural network achieves satisfactory results by using the generalization ability of the independent sample evaluation model.The R²of the whole plant and aboveground biomass models are higher than 0.63,and have smaller r RMSE and TRE.In this study,the complex relationship between satellite remote sensing data and biomass was effectively established by comprehensive application of air-space-ground monitoring method,taking airborne LiDAR as a link,combing the ground survey data and satellite remote sensing data closely,and introducing the deep learning method of cyclic neural network.This method provides a high precision continuous monitoring method forforest biomass estimation at different scales,such as single tree,stand,management unit and macro-region It is a useful exploration for obtaining forest biomass quickly and accurately in a large scale.