Accuracy Analysis And Improvement Methods For Forest Structure And Functioning Parameters Of GEDI Products

Global Ecosystem Dynamics Investigation(GEDI)is the first spaceborne high-resolution large footprint waveform laser ranging system,which is specifically designed for the global measurement of forest structure and functioning.The released forest parameter products of GEDI including forest height,leaf area index,and biomass provide new opportunities to innovate the way of research of global terrestrial ecosystem carbon cycle.Comprehensive and quantitative accuracy analysis of GEDI forest parameter products is a prerequisite for applications of GEDI products,but such studies are still lacking.Moreover,there are also limitations in methods used for producing GEDI forest parameter products.Specifically,forest height mapping based on the GEDI forest height product does not fully utilize dense GEDI observations,and the mapping accuracy is suffered from the effect of the saturation of remote sensing data on forest height.The interaction equation between the laser pulse and forest based on the Geometric Optical and Radical Transfer(GORT)to produce GEDI leaf area index product ignores the multiple scattering resulting in the estimation bias.GEDI biomass products do not fully explore the vertical and horizontal structural information of vegetation contained in GEDI waveforms and solely rely on forest height for the estimation leading to insufficient explanatory power for estimating biomass.For an appropriate application of GEDI products,it is urgent to comprehensively quantify the accuracy and develop improved methods for GEDI forest parameter products across various environmental gradient and forest types.To the best of our knowledge,this is the first study to comprehensively evaluate the performance of multiple GEDI products and analyzes the direct and indirect impact degree of multiple factors to the accuracy of GEDI products taking large amount of airborne Li DAR datasets with consistent observation time,location,and physical meaning with GEDI observations over a large scale as the reference.Meanwhile,improved methods for forest height mapping,leaf area index estimation,and biomass estimation were developed.The main works and conclusions of this study are as follows:(1)Accuracy analysis and improved mapping method of GEDI forest height productAccuracy evaluation was conducted on GEDI forest height product taking field-measured forest height collected in the Fenghuang Mountains and the National Ecological Observatory Network(NEON)Li DAR collected over 33 sites as reference.The Lindeman-Merenda Gold(LMG)was used to calculate the variance contributions of each factor to the GEDI performance of forest height measurement.Results showed that GEDI underestimated forest height(Bias:-0.78m).The high bias of GEDI forest height observation was mainly distributed in dense forests with complex terrain.Among various factors,the number of peaks,coefficient of variation of height,and vegetation cover were the three most important factors affecting the accuracy of the GEDI forest height product.The impact of geographical location error of GEDI varied significantly among different forest types.The difference in RMSE between with-and without geolocation correction ranged from 1.44 m to 5.74 m.Factors have the potential to act as a filter to screen high-quality GEDI observations,but that will cause various degrees of data loss.A combination usage of multiple factors for data screening can result in a data loss of nearly 80%of GEDI observations,namely GEDI auxiliary observations.A point-surface fusion method for forest height mapping(FPSF-CH)was proposed.The method generated an antecedent wall-to-wall forest height map which was then processed with a calibration procedure.Specifically,the random forest model was employed as a sub-model in the FPSF-CH for the generation of an antecedent forest height map(RF-CH)by integrating remote sensing data and high-quality GEDI forest height observations.For the calibration,a weighted linear regression model was employed to correct the RF-CH using nearest neighbor GEDI auxiliary observations weighed by their quality.With the full utilization of GEDI auxiliary observations,the FPSF-CH can mitigate the effect of the saturation of remote sensing data on forest height to improve forest height mapping accuracy.The experimental results showed that FPSF-CH has lower RMSE(3.48～4.02 m)and%RMSE(0.14～0.19)values compared to a global forest height product(RMSE:5.1～6.3 m and%RMSE:22.71～25.06).Meanwhile,the saturation threshold of forest height can be increased by nearly 3 m compared to the RF-CH.(2)Accuracy analysis and improved method of GEDI leaf area index productAccuracy evaluation were conducted on the GEDI effective leaf area index product(GEDI LAIe)taking field-measured LAIe and the NEON Li DAR LAIe collected over 33 sites as reference.For effect factors analysis,a method of combination Lindeman-Merenda Gold(LMG)with structural equation model was employed to quantify the relative importance and effect of each factor on the accuracy of GEDI LAIe.Results showed that GEDI underestimated values of the leaf area index,with the Bias range of-1.91～-0.02 m~2/m~2.Among forest types,dense broadleaf forests and mixed forests have higher errors than coniferous forests.Among different factors,gap fraction,vegetation characteristics,and soil characteristics are the main factors affecting the accuracy of the GEDI leaf area index product.Vegetation and soil characteristics have both direct and indirect effects on the accuracy of the GEDI leaf area index product,among which the indirect influence takes the gap fraction as the pivot.This study proposed an improved GORT method for leaf area index estimation which considers the multiple scattering of GEDI waveform,named m GORT.The m GORT improves the theoretical expression of the interaction equation between the laser pulse and forest by considering multiple scattering of GEDI pulse at ground and vegetation which were formulated by the multiple bounce theory and the unified model of vegetation bidirectional reflection.The gap fraction for leaf area index estimation was then formulated based on the m GORT.The effectiveness of the m GORT for improving the accuracy of leaf area index estimation was validated using the experiment with simulated datasets of DART and GEDI observations under different scenarios.Compared to the GEDI leaf area index product,the RMSE and%RMSE of leaf area index estimation based on m GORT can be reduced by 1.25 m2/m2 and 0.23,0.57 m2/m2 and 0.11,taking simulated data and GEDI observations as a reference,respectively.(3)Accuracy analysis and improved method of GEDI biomass productAccuracy evaluation was conducted on GEDI biomass products taking both field-measured biomass and the NEON biomass datasets collected over 19 sites as reference.The LMG and structural equation model were used to quantify the relative importance and effect of each factor on the accuracy of GEDI biomass product.Results showed that the GEDI biomass product had a moderate relative error(%RMSE:0.19～0.5).Among forests,coniferous and mixed forests with low biomass had higher relative errors(%RMSE:0.56～0.63)than coniferous forests and broadleaf forests with high biomass(%RMSE:0.23～0.33).Among various factors,the simulated waveform strategy deviation,vegetation heterogeneity,and vegetation characteristics are the main factors that affect the accuracy of GEDI biomass products.Diverse effect factors indirectly affect the error of GEDI biomass product,taking the simulated waveform strategy deviation as the pivot.A multi-stage forest structure-and waveform-driven method for GEDI above-ground biomass estimation(MTDL-AGBD)was developed.This method utilizes airborne Li DAR to estimate biomass as training data for GEDI biomass estimation to avoid the negative impact of the simulated waveform strategy deviation.Moreover,a multi-task deep learning framework was developed,which consists of two deep learning tasks.Specifically,the main task takes forest structure parameters as inputs,while the auxiliary task takes GEDI waveform as inputs to dig the vertical and horizontal structural information of the forest,providing supplementary information for the main task.The experimental results showed that compared to the GEDI biomass product,the%Bias,%MAE,and%RMSE of MTDL-AGBD decreased by 0.60,0.58,and 0.54,respectively.The comprehensive analysis of GEDI products’accuracy in this study can answer if the GEDI forest parameter products meet a specific data quality requirement of the carbon cycle and ecosystem process models.The studies on methods for improving forest parameter estimation can provide a reference for improving product accuracy and promote the quantitative application level of GEDI.