Research On Inversion Algorithm Of Forest AGB Machine Learning For SAR Data In Saihanba Area

Forests are important transit points for carbon sinks and water cycles,storing 50 to 65percent of terrestrial organic carbon and contributing half of terrestrial productivity.The major problem of forestry carbon sinks is how to continuously and accurately calculate the above-ground biomass of forests in large-scale areas.The emergence of remote sensing technology provides a solution to this problem.Optical remote sensing can only collect the information of electromagnetic waves reflected from the upper surface of the forest.Microwave remote sensing can penetrate the forest canopy and collect more information on the vertical structure of the forest.Therefore,the use of synthetic aperture radar has a unique advantage in the monitoring of above-ground forest biomass.Based on the ALOS-2 full-polarization high-resolution synthetic aperture radar data,the research performs radiometric calibration,multi-look,filtering,terrain correction and geocoding processing on the image data to obtain accurate backscatter coefficients and polarization decomposition parameters.Use secondary survey data to remove non-forest areas from the imagery.Taking Saihanba Mechanical Forest Farm in Chengde City,Hebei Province as the research area,82 sample plots were measured on the ground,and 12 kinds of polarizations based on physical models,such as Freeman three-component decomposition,NEED decomposition and MCSM five-component decomposition,were explored by pixel statistics method.The total scattering power and the proportion of each scattering component of the decomposition method in the forest area were analyzed qualitatively to determine the dominant scattering mechanism of the forest,and quantitative analysis was used to explore the changes of different polarization decomposition methods on the same scattering mechanism.On this basis,correlation was used.Analytical methods The pros and cons of each polarization decomposition method and the progressive relationship were analyzed to find the most suitable polarization decomposition method for modeling and estimating forest above-ground biomass.In order to solve the problem of estimating forest aboveground biomass from fully polarized data with few sample plots,large number of independent variables and high collinearity,this paper proposes a Random Forest Mixture Adaptive Genetic Algorithm(RF-AGA).At the algorithm level,the problem of high collinear regression estimation and poor generalization ability between the polarization decomposition parameters is solved.The main research conclusions are as follows:(1)Generally speaking,each polarization decomposition method basically satisfies the stability of the total scattered power,and in the forest area,each polarization decomposition method basically satisfies the total power unchanged,and the proportion of volume scattering in the forest area is significantly higher than that in the forest area.Among other scattering variables,the L-band lower body scattering still dominates.(2)Compared with the Freeman three-component decomposition,the VanZyl decomposition in the L-band decreased by 32.4%in the forest area,and the decrease was larger than the theoretical deviation,but it still failed to affect the qualitative explanation of the forest-dominated scattering mechanism.(3)VanZyl three-component decomposition,Yamaguchi three-component decomposition and Freeman two-component decomposition are most suitable for estimating forest aboveground biomass.(4)Huynen decomposition and Barnes decomposition are completely equivalent to D1,D2,D3 assumptions of Unified Huynen decomposition,and the estimation of forest aboveground biomass is most suitable for D7 assumption.(5)The RF-AGA algorithm and the stepwise regression method are simultaneously applied to the high-dimensional SAR dataset of 82 samples and 340 independent variables.The test accuracy of RF-AGA algorithm is R~2=0.767,RMSE=16.54t/hm~2,rRMSE=20.37%,and the test set accuracy of stepwise regression method is R~2=0.335,RMSE=26.62t/hm~2,rRMSE=31.09%.In the case of collinear SAR data sets,the accuracy of RF-AGA algorithm is significantly better than that of stepwise regression.