A Study On The Estimating Method Of Forest Above Ground Biomass Based On Remote Sensing Data

Forest biomass is the basal parameter to characterize forest ecological functions and evaluate its values. It is an essential component in forest carbon storage and global carbon cycle researching. Accurately estimate the forest biomass has important significance in mastering the regional and global carbon storage and distributions. Traditional forestry inventory has long period time, can only acquire the plot data in the cost of forest destructive. Remote sensing technology, as an effective means, can acquire large-scale continuous data. Optical remote sensing data has been widely used in forest parameters modeling, it can get the forestry horizon information but has difficulty in acquiring vertical information. The optical remote sensing data is easy to have spectral signal saturation and poor penetration. Laser scanning technology, with its high accuracy to obtain forestry vertical structural information, is rapidly developed in forestry application researches.Forest biomass estimation methods include traditional parametric and non-parametric machine learning algorithms. Traditional method present the model equation, in contrary non-parametric method has a problem of secret operations, using the algorithm and output model calculation result directly. In order to meet the non-linear relationship between remote sensing variables and forest biomass, the non-parametric machine learning algorithms develop rapidly in the application of forestry research.Based on the airborne LiDAR point cloud data, Landsat-TM5 data, HJ1B-CCD2 and field inventory data of Genhe Chaocha Forest Farm in Inner Mongolia, sets the cold temperate primary and secondary forests as research subjects. Generating model and mapping of forest canopy height, forest canopy closure and forest AGB. Investigate the applicability of double tangent tree crown recognition algorithm and find out the difference between Lorey’s height and crown area weighted height. In finally, select the optimal forest height model and calculate the spatial distribution of forest height and provide reference data for subsequent studies of biomass and carbon storage research. Using LiDAR density variables and field measured canopy closure to generate canopy closure model. Combine TM, HJ and LiDAR data, using multiple stepwise regression, random forests algorithm (RF), support vector regression algorithm (SVR) and Maximum entropy (MaxEnt) methodologies to estimate forest AGB. Make full use of information from multisource remote sensing data and select the optimal modeling approach to estimate the forest AGB.The main conclusions are as follows:(1) LiDAR 50% percentile height has a strong significant correlation with field measured height and can explain more variation of field measured height. LiDAR optimal height model by LiDAR 50% percentile height and field measured Lorey’s height, the model R2 is 0.869, RMSE is 1.36m and its validate average estimate accuracy is 94.73%. The LiDAR crown area weighted height, was calculated by double tangent tree crown recognition algorithm, combining with field Lorey’s height to generate forest height model. LiDAR 50% percentile height model has higher accuracy than LiDAR crown area weighted height model. The mixed forest estimated accuracy is higher than the coniferous forests.(2) LiDAR density variables represents well of the overall forest canopy closure and the model R2 is 0.784, RMSE is 0.077, the testing accuracy is 88.29%, mixed forest estimation accuracy is higher than the coniferous forest.(3) The study of forest AGB showed:multiple linear stepwise regression result of LiDAR data:model training R2 is 0.69, testing RMSE is 23.09 t.hm-2 and the average estimated accuracy is 82.51%. After inputting forest height and forest canopy closure, forest AGB model estimated accuracy of TM data and HJ data have significantly improved. Random forest regression algorithm on forest AGB:LiDAR data has the highest accuracy, the model R2 is 0.835, RMSE is 18.264 t.hm-2, testing RMSE is 20.138 t.hm-2 and the average estimated accuracy is 91.359%. After inputting forest height and forest canopy closure, the forest AGB model estimated accuracy of TM data and HJ data has significant improved. The SVR algorithm estimated results:LiDAR data model training R2 is 0.854, RMSE is 17.557 t.hm-2, testing RMSE is 19.004 t.hm-2 and the average estimated accuracy is 80.374%. After inputting forest height and forest canopy closure, the forest AGB model estimated accuracy of TM data and HJ data has significant improved. Maximum Entropy model algorithm estimated results:LiDAR data has higher estimated accuracy in forest AGB modeling than other remote sensing data. Modeling with HJ data, TM data, forest height and forest canopy closure, the estimated accuracy of forest AGB model is better than other data combinations’ results.(4) RF feature select and fselect feature select algorithm improved the model estimated accuracy. The RF feature select algorithm has significant improving effects.(5) LiDAR data have good mathematical correspondence in four algorithms. Its estimation results are stable and better than other remote sensing data. RF regression algorithm has strong learning ability and model training accuracy. SVM has strong capability in processing small sample training data. MaxEnt algorithm has higher estimated accuracy on high biomass. Forest height has obviously effects on improving the model accuracy than forest canopy closure. The forest canopy closure has slightly effects on improving the forest AGB estimated accuracy and the impact is not significant.