Estimation of Forest Aboveground Biomass in Changbai Mountain Region Using ICESat/GLAS and Landsat/TM Data

Hong Chi, Guoqing Sun, Jinliang Huang, Rendong Li, Xianyou Ren, Wenjian Ni, Anmin Fu
2017 Remote Sensing  
Mapping the magnitude and spatial distribution of forest aboveground biomass (AGB, in Mg·ha −1 ) is crucial to improve our understanding of the terrestrial carbon cycle. Landsat/TM (Thematic Mapper) and ICESat/GLAS (Ice, Cloud, and land Elevation Satellite, Geoscience Laser Altimeter System) data were integrated to estimate the AGB in the Changbai Mountain area. Firstly, four forest types were delineated according to TM data classification. Secondly, different models for prediction of the AGB
more » ... iction of the AGB at the GLAS footprint level were developed from GLAS waveform metrics and the AGB was derived from field observations using multiple stepwise regression. Lastly, GLAS-derived AGB, in combination with vegetation indices, leaf area index (LAI), canopy closure, and digital elevation model (DEM), were used to drive a data fusion model based on the random forest approach for extrapolating the GLAS footprint AGB to a continuous AGB map. The classification result showed that the Changbai Mountain region was characterized as forest-rich in altitudinal vegetation zones. The contribution of remote sensing variables in modeling the AGB was evaluated. Vegetation index metrics account for large amount of contribution in AGB ranges <150 Mg·ha −1 , while canopy closure has the largest contribution in AGB ranges ≥150 Mg·ha −1 . Our study revealed that spatial information from two sensors and DEM could be combined to estimate the AGB with an R 2 of 0.72 and an RMSE of 25.24 Mg·ha −1 in validation at stand level (size varied from~0.3 ha tõ 3 ha). Remote Sens. 2017, 9, 707 2 of 25 forest inventories; these methods can obtain good AGB estimation. National inventories (e.g., National Forest Resource Inventory Program of the State Forestry Administration, China, or the U.S. Forest Service Inventory and Analysis Program) provided the most detailed field estimates of forest AGB at the national scale [6, 7] . However, the lack of field measurements in remote areas and the inconsistency of data requirements among different administrative units are the major constraints to obtain sufficient regional AGB estimation using field-based methods [5, 8] . In addition, obtaining comprehensive, spatially-complete, temporally uniform, and accurate forest inventory data is usually time-consuming and labor-intensive over very large areas. Reducing the uncertainty in the AGB estimates requires spatially-continuous observations that are fine enough to capture the variability over a landscape that may undergo natural disturbances or landuse changes [5, 9] . Remote sensing has the capability to map forest AGB over wide geographical areas. However, as no remote sensor has been developed that is capable of providing direct measurement of AGB, additional field-sampled AGB is required to correlate with canopy reflectance measured by passive optical sensors or backscatter intensity from Synthetic Aperture Radar (SAR) sensors from moderate to fine spatial resolution [10] . Optical sensor data are appropriate for the retrieval of forest horizontal structures, such as forest types and canopy cover, due to its spectral sensitivity to different species [11] . Furthermore, satellite-based optical sensors are perceptive to other forest structural variables, such as species abundance [12], basal area [13], stem density [13] , and crown size [14] , which are correlated to some degree with AGB because forest spectral reflectance contains information on the vegetation chlorophyll absorption bands in the visible region and sustained high reflectance in the near-infrared region [15] . Previous studies have demonstrated the sensitivity of visible and shortwave infrared wavelengths to AGB [4, [16] [17] [18] . In addition to the usage of single-band spectral signatures, various vegetation indices (VIs) derived from TM (Thematic Mapper), ETM+ (Enhanced Thematic Mapper Plus), MODIS (Moderate Resolution Imaging Spectroradiometer), etc., as primary data sources have been used to estimate forest AGB in different regions, such as tropical America and Asia, Central Europe, East Africa, the U.S., and China [16, [18] [19] [20] [21] . Moreover, some optical sensor data, such as IKONOS, ALOS/PRISM (Advanced Land Observation Satellite/Panchromatic Remote-sensing Instrument for Stereo Mapping), and ZY-3 (Zi Yuan 3), provide a stereo-imaging capability that can be used to predict forest canopy structure [22] [23] [24] . Although passive optical remote sensing is widely used for AGB estimation, frequent cloud coverage in mountains and moist regions, and the low saturation level of spectral signatures on AGB estimation in medium to high AGB forest (i.e., the spectral reflectance or vegetation index of optical remote-sensing sensors are of limited value in medium to high biomass forests) are the major disadvantages of this technology [11, 25, 26] . Recently, much attention has been paid to the use of LiDAR (Light Detection And Ranging) remote sensing in AGB estimation due to its potential to capture the vertical structure of vegetation in great detail and it is the only sensor type available at present whose signal does not saturate in high AGB forests (e.g., 1200 Mg·ha −1 and 1300 Mg·ha −1 ) [27] [28] [29] [30] . LiDAR footprints may be used as a manner of sampling similar to field plots, wherein the derived AGB is used to integrate optical remotely-sensed data in order to facilitate stratification, thereby extending plot-level AGB to large areas [31] . Airborne LiDAR provides highly accurate AGB estimates [32] [33] [34] , but the associated large data volume, as well as the sophisticated technical equipment and high acquisition costs to observe remote areas, usually limit its application at regional and global scales [16] . The Geoscience Laser Altimeter System (GLAS) sensor onboard the Ice, Cloud, and land Elevation satellite(ICESAT) satellite provided data freely and has proven itself suitable for AGB and canopy height estimation over continental-to-global areas [3, 4, [35] [36] [37] . One major limitation of GLAS was the lack of imaging capability and the fact that it provided relatively sparse sampling information on forest structures [38] . Therefore, GLAS data are often integrated with imaging optical systems to estimate forest structural variables at the regional scale [39, 40] . Due to the concerns of global climate change, as well as large swaths and daily availability of moderate spatial resolution sensors data (e.g., MODIS), AGB estimation that integrated LiDAR with Remote Sens. 2017, 9, 707 3 of 25 MODIS over continental-to-regional scales has gained increasing attention in the last few years [3, 4, 8, 41] . However, moderate spatial resolution data loses spatial detail of AGB variability, and it is often difficult to associate field data with satellite observations because the inconsistencies in spatial resolutions result in a mismatch between filed plots and remotely-sensed imagery [11, 42] . While medium resolution data, such as TM/ETM+, provides spatial detail compatible with the size of vegetation units and AGB field measurements [16] , there are increasing demands and attempts for estimating AGB at finer resolution from medium-resolution sensors [43] . The integration of Landsat and GLAS data for large-area applications has become more practical with the free availability of their archives and the capability of near-global coverage [44] . Duncanson et al. [44] combined Landsat/TM and spaceborne LiDAR to estimate AGB in South-central Canada, and found that the data integration is most useful for forests with an AGB less than 120 Mg·ha −1 , an age less than 60 years, and a canopy cover less than 60%. However, a generic GLAS-based AGB model was developed from field plot data and GLAS waveform metrics, regardless of forest types. Moreover, an AGB map generated from the model that was developed from field measurements and a serial of airborne LiDAR metrics was designed to provide a reference for validation of AGB estimates. It was inevitable to introduce potential errors when considering additional processing procedures of airborne LiDAR and the mismatch between airborne LiDAR-derived AGB and GLAS-derived AGB in spatial resolution. Zhang et al. [5] presented a simple parametric model that integrated leaf area index (LAI) estimates from Landsat and canopy height from GLAS for conifer-dominant forests of California. The relationships between the GLAS-derived maximum canopy height and Landsat-derived LAI were modeled using a linear model, which was based on the assumption that the power law between LAI and the maximum tree height is a first order approximation. In fact, the relationship between the two parameters was not linear, which would introduce uncertainty into the estimation of the maximum tree height, which was used for AGB estimation. Li et al. [45] described a Landsat-lidar fusion approach for modeling canopy heights of young forests by integrating historical Landsat imagery with GLAS data. They used two methods to explore the relationships between forest height and predictor, including stepwise linear regression (SLR) and regression tree (RT). The RT models yielded substantially lower RMSD (root mean square of the difference) than the SLR models when use three different variables groups. The best RT model was developed when forest age metrics, NDVI metric, NBRI (normalized burn ratio index) metrics and IFZ (integrated forest z-score) metrics (that is an index as a measure of forest likelihood) were used together. Although RT can use multiple linear equations to approximate nonlinear relationships, the substance of the final models in this study had only two linear regression equations. Obviously, the linear regression models will show more overfitting with the increase of independent variables. Dolan et al. [46] combined forest age information derived from Landsat data with canopy height data from GLAS to quantify rates of forest growth in the eastern United States. They regressed the GLAS-derived heights against the age of last disturbance in linear models to yield vertical growth rates. The growth rates that combined with height-biomass allometric relations can be converted to estimates of AGB. Due to lack of field measurements of AGB co-located with GLAS footprints, they used age (or growth rates) as an intermediate variable to convert GLAS-derived height to AGB at landscape-scale (the first step: GLAS-derived height to AGB at footprint-scale; the second step: linking relationships between GLAS-derived AGB and Landsat-derived age; and, the third step: extrapolating GLAS-derived AGB to landscape-scale using the relationships in the second step). The used allometric relations between AGB and GLAS-derive height was derived from mixed forest region. Helmer et al. [47] used a method similar with Dolan et al. [46] to estimate AGB of old-growth in Brazil. They also used age as an intermediate variable to convert GLAS-derived heights to AGB at landscape-scale. Be different from the other research, the used allometric relations between AGB and GLAS-derived height was derived from evergreen broadleaf forest region. In summary, to the best of our knowledge, few published studies have explored forest-type-specific AGB estimation integrated spaceborne LiDAR and Landsat data at the regional level. In these studies, both the forest-type-specific prediction models of AGB and
doi:10.3390/rs9070707 fatcat:vxlegf26kvbznbu6iegr53azo4