Lidar Position Calibration And Biomass Estimation Based On Deep Neural Network

As the main body of terrestrial ecosystems,forest biomass accounts for more than 80% of the global terrestrial ecosystem biomass,and plays a vital role in mitigating the impact of global climate change,maintaining biodiversity and preventing soil erosion.Forest aboveground biomass is a key parameter for evaluating forest productivity and carbon sequestration rate.Therefore,quickly and accurately estimating forest aboveground biomass is very important for quantifying carbon storage and understanding the global carbon cycle.China plans to launch a terrestrial ecosystem carbon monitoring satellite by the end of 2020.This is China’s first satellite specifically for the application in forest monitoring.The satellite will be equipped with a large-footprint Lidar system,which can be used to the research of forest vertical structure detection,inversion of average forest height,forest biomass,etc.When retrieving forest structure parameters using large-footprint lidar data,the forest height at the spot scale changes rapidly,and often requires joint optical image data,which places high requirements on the horizontal positioning accuracy of the Lidar’s footprint.In this paper,the use of Lidar data obtained by the Academy of Inventory and Planning of National Forestry and Grassland Administration in Zhangjiakou,Hunan and Zhangjiajie,Hebei is used to study the position calibration of large-footprint Lidar and the estimation of forest above-ground biomass based on the simulated large-footprint Lidar waveform.The main research contents of this article include:1.Using laser point cloud data to generate Leaf Area Index(LAI)and Canopy Height Model(CHM).The comparison with Digital Orthophoto Map(DOM)shows that the distribution of vegetation in LAI and CHM is consistent with that in DOM.The linear regression model is established by using the ground sample survey data and the product of LAI and CHM,so that the forest ground biomass sample data is obtained,which greatly expands the number of training set samples of the biomass estimation model.2.The deep metric learning model is used to measure the similarity of the echo waveform of the large-footprint Lidar.The results show that the Multi-Layer Perceptron(MLP)has a better measurement effect,and the Long Short Term Memory(LSTM)has the least overfitting,while the one-dimensional Convolutional Neural Network(CNN)is somewhere in between.At the same time,by modifying the distance metric function and adding a sequential neural network after the neural network,a deep metric learning model with better performance can be obtained.3.The waveform parameters of the simulated lidar waveform are used toestimate the aboveground biomass of the forest.There are 16 waveform parameters in total,and 22 different parameter combinations are constructed.A MLP biomass estimation model with 5 hidden layers was designed and compared with the estimation results of multiple linear regression models.The results show that among the 21 parameter combinations,the MLP model has achieved a significantly better estimation effect than the multiple linear regression model.Compared with the measured biomass data of the sample site,it is found that the estimated results of the two models are generally lower than the measured value of the sample site,but the deviation of the MLP model is significantly smaller than that of the multiple linear regression model.Experimental analysis confirmed the advantages of the deep neural network model in footprint position calibration of lidar and estimation of forest aboveground biomass based on lidar data.