Forest Parameters Inversion At Different Spatial Scales Using Multi-source Remote Sensing Data

As the largest and most important natural ecosystem in the terrestrial ecosystem,forest play a crucial role in maintaining global ecological balance and promoting global biological evolution and community succession.Forest parameters,as the most basic quantitative representation of forest ecosystems,can reflect the relationship between material circulation and energy flow between forest and the environment,and are important indicators for assessing forest carbon sources and carbon sinks.Forest canopy height and above-ground biomass are two important forest parameters.The quantification of its regional distribution can play an important theoretical and practical significance for the study of forest ecosystems.Traditional forest canopy height and above-ground biomass measurements are mainly based on artificial field surveys.This method is time consuming and difficult to apply to large-scale forest surveys.Remote sensing,as a new discipline developed at the end of the 20th century,is providing new research methods for the quantitative estimation of forest parameters from a unique observation perspective.The optical remote sensing image has spatial distribution continuity and can provide diversified spectral information,and can provide corresponding remote sensing feature expansion factors for inversion of forest parameters at the regional scale.However,the signal penetrating ability is poor and only 2D plane spectral information of the measured object can be obtained.As an active remote sensing technology developed rapidly in recent years,LiDAR has strong penetrating ability for forest canopy and can obtain vertical distribution of forest structure.The airborne LiDAR can observe ground objects in a top-down scanning manner.The three-dimensional LiDAR point cloud can accurately describe the three-dimensional structure characteristics of forests and is widely used in forest-scale parameter estimation of forest stands.However,its data acquisition costs are high and are easily limited by weather conditions,making it difficult to cover large areas.The Geosciences Laser Altimeter System(GLAS),because of its laser sensor being placed on the Ice,Cloud,and land Elevation Satellite(ICESat),emits 1064-nm laser pulses and receives the returned laser energy from a variable diameter~70m footprint.It globally provides full waveform Li DAR data for the representative sample of footprints spaced at~170 m.Due to the large footprint property of GLAS,in the area where the topography changes greatly,the waveform shape within the footprint range is easily affected by the topography,resulting in distortion,which makes the forest parameter estimation is biased at footprint scale.Therefore,how to make full use of the advantages of different remote sensing data sources,the collaborative inversion of forest parameters at different spatial scales(footprint scale,forest stand scale and regional scale)becomes the focus of this study.On the basis of other scholars'research,the Greater Khingan Range,a major forest-covered area in northeastern China,was used as an example to carry out a method for the collaborative inversion of forest parameters based on multi-source remote sensing data on different spatial scales.The paper research content includes the following three aspects:1)Inversion of forest parameters at forest stand scale.Using airborne LiDAR point cloud data estimate stand forest canopy height;using fusion of airborne LiDAR features and landsat TM image texture features estimate stand aboveground biomass.2)Inversion of forest parameters at footprint scale.Using the satellite GLAS waveform parameters and SRTM terrain data establish a terrain correction model to invert GLAS_based forest canopy height;using GLAS_based forest canopy height and GLAS waveform parameters invert GLAS_based forest aboveground biomass.3)Inversion of forest parameters at regional scale.GLAS_based forest parameter inversion results are used as training samples.MODIS standard products are used as extension factors.FAO ecological zones data is modeled zoning standard.SVR model is used as spatial modeling method to invert regional scale forest parameters.The mainly findings are as follows:1)By analyzing the distribution characteristics of airborne LiDAR point cloud in coniferous forests,broad-leaved forests,and mixed forests,it can be concluded that the point cloud canopy profile of coniferous forests changes more clearly,and the distribution of laser points under the crown and canopy is relatively obvious less.Broad-leaved forests can capture more laser points due to the larger crown area,which makes the canopy delamination of the point cloud lessen.Because the shrubs under mixed forests are more complex and there are no obvious signs of disturbance,there are more laser points at a height of 0-5m than other pure forests and the canopy cloud profile changes more gently.Therefore,this clear canopy cloud distribution feature makes the point cloud characteristics of airborne LiDAR accurately reflect the forest canopy height.2)According to the height distribution of the point cloud,it is concluded that the CHM generated by the airborne LiDAR point cloud has a better response to the forest canopy height.Due to the hollow effect of CHM,the improved smooth Gaussian convolution filtering method is used to smooth the CHM and achieve a better smoothing effect.Lorey's high and arithmetic mean heights are used to verify the accuracy of the smoothed CHM.Lorey's height are closer with CHM,and the verification accuracy of broadleaf forest is the highest,R~2=0.74,RMSE=1.14m,G=93%.3)By comparing the modeling results of the texture features of optical image,the point cloud features,and integrated image texture features and airborne LiDAR point cloud features,found that the green band reflected weaker differences in the forest aboveground biomass,and the modeling accuracy was lower.The modeling accuracy of 7×7 window in red band was the highest,R~2=0.72,RMSE=20.53t/ha.Larger window size made the difference between pixels amplified,and resulting in higher modeling accuracy.The precision of the quantile of point cloud height was in a normal distribution trend as the quantile rises,and the H50 had the highest modeling accuracy,R~2=0.74,RMSE=20.01t/ha.The modeling accuracy of the fused image texture features and point cloud features has been improved to a certain extent compared with single variables,and the remote sensing factors retained by different forest type models were quite different.Mixed forest modeling accuracy was highest of R~2=0.76,RMSE=20.13t/ha.4)GLAS waveforms will have a broadening effect with the influence of topography.By introducing the terrain slope and footprint diameter,an improved terrain calibration method was proposed for GLAS_based forest canopy height estimation.The improved terrain calibration method can reduce the influence of slope.The RMSE of different grades was stable between 3.26 and 3.88m.The verification accuracy of the broad-leaved GLAS_based canopy height was the highest.Its verification RMSE and verification accuracy G were 1.80m and 89%,respectively.5)Comprehensive utilization of factors such as GLAS_based forest canopy height,GLAS waveform trailing length,Waveform leading length,and waveform length,estimated GLAS_based forest aboveground biomass combined with multiple linear regression models.The modeling R~2 and RMSE did not decrease with the increase of slope grade,and the verifing accuracy and modeling accuracy kept a good consistency.Compared with the verification accuracy of the aboveground biomass of forest stands,the verifiing accuracy of different forest types had a similar trend,but the GLAS_based forest aboveground biomass verifing accuracy has been reduced.6)Using machine learning support vector regression algorithm and combing with GLAS_based forest canopy height result and MODIS remote sensing features,SVR modeling were built for forest canopy height at regional scale.There were great differences in the values of the optimal model parameters for different kernel functions.The modeling accuracy and time complexity of the RBF kernel function are better than those of other kernel functions.Because of the differences in the distribution range of ecological zones,the MODIS remote sensing features of some ecological zones can not accurately reflect the forest growth status of this sub-region,resulting in a large difference between modeling R~2 and MSE in different ecological zone.In addition to the Ba ecological zone,the estimating RMSE of SVR models in other ecological zone was relatively small comparing with the multiple linear regression models.Using different types of verification data to verify the regional canopy height estimation results,there was a good consistency between the verification results.7)Combined with MODIS features and regional forest canopy height result,GLAS_based forest aboveground biomass was used as a training sample to conduct SVR modeling of regional scale.Compared with only using MODIS features modeling results,the introduction of regional forest canopy height variables has improved the model fitting degree,and the modeling accuracy of coniferous forest and mixed forest model has increased more obvious.Compared with the forest canopy height model at the regional scale,the aboveground biomass models of coniferous forests,broad-leaved forests,and mixed forests are consistent in modeling accuracy in different ecological zones.