Estimation Of Forest Aboveground Biomass By Integrating ICESat/GLAS Waveform And TM Data

Forest biomass is an important indicator for assessing the ecosystem productivity,and also the basis for analysis of substance circulation in forest ecosystem. As themain body of the Earth’s terrestrial biosphere, forest plays a major role in fixingatmospheric CO2and mitigating climate change. The traditional methods forcalculating biomass rely on a considerable amount of in-situ measurements thatinvolves extreme time and labour, which is also difficult to update spatial distributionin a large area. However, with the rapid development of remote sensing technique,multi-source remote sensing data including aerial photographs, optical images, radarand LiDAR have been extensively used for monitoring the forest types, spatialdistribution and structural features, which provides a fast, cheap, and convenient wayto estimate the large-scale forest biomass and long-term dynamic change. Usingoptical remote sensing data to retrieve regional forest structural parameters andbiomass started earlier. However, owing to weak penetrating power, it mainly reflectshorizontal forest structures. Synthetic Aperture Radar (SAR) could penetrate thevegetation canopies to a certain extent, but it was seriously disturbed by thetopographic relief and no longer sensitive when the crown was closed or the biomasswas very high, which restricted its application in regional estimation. LiDAR (LightDetect and Ranging) is an active remote sensing that developed rapidly in recent years,and has performed great potential and advantage in estimating forest vertical structureparameters due to its strong capacity to penetrate forest canopies. The small-footprintLiDAR can provide a wide variety of efficient stand parameters with high precision,but it still has plenty of limitations such as high cost, limited coverage, massive amounts of data, and so on. While large-footprint ICESat/GLAS waveform data canestimate spatial structure parameters of large-area forest, and acquire more standinformation than small-foortprint LiDAR through recording the time variation in theintensities of returned laser pulses, which resole elliptical areas approximately65m indiameter. Futhermore, the satellite coverage is even global scale in theory. CurrentlyICESat/GLAS data have being widely used in estimation of ecological parameters offorest. However, the major shortcoming of the large-footprint LiDAR is lack ofimaging capability, which could only provide vegetation samples in spatially discretefootprints and restrict its application in regional biomass estimation. The developmentof multi-source remote sensing techniques stimulates how to effectively use these datafor ecological studies owing to limited information from a single sensor. At present,there is a trend of using multi-sensor remote-sensing system integration and datafusion (such as GLAS waveform data and optical images or SAR), which indicatesgreat potential in forest biomass estimation.In view of the complicated terrain in Changbai Mountains, the study explored thefeasibility of extraction of forest canopy height and above-ground biomass byintegrating multispectral data (TM) and GLAS waveform data. The algorithms ofprocessing the GLAS data were implemented and the models of estimating canopyheight in the area of flat slope and steep slope were established. Considering theGLAS waveform data were spatially discrete, the study brought forward theestimation model of regional forest canopy height using optical remote sensingimages. At last, the forest above-ground biomass in the study area was retrieved byintegrating the canopy height from GLAS with optical data. The results showed thatTM wide-swath data and GLAS waveform data can be combined to estimate forestcanopy height and above-ground biomass with good precision. The main findings areas follows:1. Based on the object-oriented method from eCognition and corrected LandsatTM data, this study acquired the land use/cover data in2010. Then the data wasfurther divided into coniferous forest, broad-leaved forest, mixed broadleaf-conifer forest and non-forestry land. Meanwhile under the ArcGIS platform, six bands ofreflectance values, ten kinds of vegetation index including RVI、NDVI、SLAVI、EVI、VII、MSR、NDVIc、BI、GVI and WI, and the topographic factors as DEM、ASPECTand SLOPE, were calculated to analyze the correlations between the correspondingforest LAI measured using TRAC with each factor. Then compared the modelperformance of multiple linear regression (MLR) and partial least squares (PLS)method, this paper established the optimal model to retrieve the forest LAI of eachforest type. At last, the distribution map of forest LAI in this study area was made byintegrating remote sensing inversion model with the forest classification data acquiredbeforehand. The results showed that it’s hard to retrieve forest LAI accurately with asingle band or vegetation index, but we could improve the model’s precision bycombining these variables. Meanwhile, the topographic factors cannot effectivelyenhance the model’s effect. From the perspective of inversion ability of each foresttype, PLS models were obviously superior to MLR. Forest crown density is animportant parameter for evaluating forest status and indicating possible managementinterventions. Based on the highly related vegetation index, this paper established theoptimal empirical model to retrieve the forest crown density of each forest type.Compared with the empirical approach, we also improved the dimidiate pixel modelby defining four-level thresholds of NDVI including2%,1%,0.5%, and0.2%minimum and maximum cut of points in the histograms of NDVI imagery. The resultsshowed that forest crown density estimation based on the dimidiate pixel model withthreshold of NDVI being0.5%was efficient with high precision over the in-situ fieldvalidation for coniferous forest and mixed forest whereas being1%for broad-leavedforest in Changbai Mountain area. And the model performed well for coniferousforest (R2of0.872and RMSE of0.019), followed by broad-leaved forest (R2of0.832and RMSE of0.016) and mixed forest (R2of0.799and RMSE of0.021).2. After analyzing the distribution of these GLAS waveform data for each foresttype in the study area, the algorithms of processing GLAS data were implemented.Currently, many studies use Gauss function to fit the GLAS full waveform, and obtain the key parameters. However, this method might lead to ground echo signals missing.Therefore, Fourier transform was proposed to filter and fit the waveform, which builta basis for extraction of the key parameters and estimation of canopy height. Thenbased on the waveform length, topographic index and centroid position, this studyestablished the estimation model of forest canopy height in complicated terrain.Overall, the precision of GLAS forest canopy height was high in flat regions (about0.5m). The RMSE in high-slope terrain also reached2.021~2.674m after correction.3. Based on the algorithm of forest canopy height for GLAS data, the models ofconifer, deciduous broadleaf forest, and mixed forest in complex terrain conditionswere established by integrating TM parameters (6spectral bands,10vegetationindexes, LAI and crown density), canopy height and the topographic factors, and thenvalidated with in-situ measured data. The results showed that the topographic factorscould supplement the shortcomings of spectral information. RMSE of the optimizedmodels for each forest type using PLS method were less than2m. Although there stillexisted overestimation or underestimation, the general trend held true.4. The relationship between forest canopy height and above-ground biomass wasanalyzed. The result showed that the correlation coefficients between above-groundbiomass and the canopy height were0.903and0.589, and the R2of the regressionmodel reached0.816and0.412, respectively, which indicated the predictive ability ofcanopy height to forest biomass. Then the multiple linear regression model and BPneutral network model of forest above-ground biomass were established bycombining forest canopy height, vegetation indexes, LAI and crown density. Theresult showed that the characteristic of nonlinearity of BP neutral network model wasmore fitted to inverse the forest biomass.