Final Report Summary - FORESTMAP (3D Forest Structure Monitoring and Mapping)
Accurate spatially-explicit information on forest structure, the 3D arrangement of forest components, is of great importance to better understand the carbon cycle, to design sustainable forest management strategies and mitigate greenhouse gas emissions from deforestation and forest degradation, as well as to halt the loss of biodiversity and to reduce the risk of wildfires. Parameters used to describe forest structure include canopy height; fractional cover (FC); canopy gap size; biomass (an Essential Climate Variable / ECV); species composition and canopy bulk density (CBD). Despite its importance, there is still limited knowledge of the spatial distribution of forest structure and its dynamics. The main goal of ForestMap was to develop an innovative and original methodology to derive 3-D information on forest structure, based on a multi-scale multi-sensor remote sensing approach. More specifically, the objectives were to: 1) develop a methodology for integrating active (LiDAR & Radar) and passive multispectral remote sensing data to estimate biomass, canopy bulk density and fractional cover and their dynamics; 2) provide spatially-explicit estimates of forest structure at regional scale based on an up-scaling approach; and 3) to provide spatial information on the uncertainty of the estimates at each scale.
The project has been carried out in two phases. The outgoing phase (01/01/2015-31/12/2016), was carried out at the NASA Jet Propulsion Laboratory, California Institute of Technology; and the returning phase (01/01/2017-31/12/2017) was carried out at the Centre for Landscape and Climate Research, University of Leicester.
The first step of the proposed methodology consisted on deriving accurate information of forest structure (aboveground biomass / AGB, canopy height, fractional cover and canopy bulk density) using airborne LiDAR data. Sensor characteristics and surveying parameters affect LiDAR measurements of the spatial distribution of the canopy components, which will impact vegetation structure metrics derived from them. Moreover, LiDAR measurements can be represented using two different data models, the echo-based or the raster model. The data model used along with the surveying characteristics will have an impact on the biophysical properties derived from airborne LiDAR data; therefore, we evaluated the ability of the metrics derived from the two data models used to represent the LiDAR information, as well as the impact of point density (number of laser hits per square meter) on the AGB estimation in three different biomes (temperate, moist tropical and Atlantic forest). In addition to this analysis, the use of multitemporal LiDAR data for forest and carbon monitoring was investigated, including tree growth and biomass / carbon dynamics of a temperate forest site.
The second step of the methodology aimed at extrapolating local airborne LiDAR estimates of forest structure to larger regions based on satellite data, both multispectral and Synthetic Aperture Radar (SAR) imagery, using a machine learning approach. First, the core of the methodology was developed to provide biomass changes caused by a megafire. Landsat-8 OLI data were used to extrapolate LiDAR-based AGB estimates over the burned area to obtain pre-fire AGB using a Support Vector Machine (SVM) regression approach. After subtracting post-fire AGB estimates it was possible to compute the amount of biomass consumed by the fire and the carbon release. This methodology was also applied to derive canopy fuel properties that are critical for fire behaviour. We also evaluated the relation between the derived properties and the burn severity.
The extrapolation of LiDAR-based forest structure estimates was further investigated using multispectral and SAR data. Thus, using LiDAR samples of canopy height over different study areas distributed across two biomes, we developed SVM regression models to extrapolate LiDAR-derived canopy height using Landsat-8 OLI and ALOS PALSAR imagery. It was found that the relative importance of the variables derived from sensors in the model performance varied depending on forest structure and composition. Thus, whereas over the temperate broadleaf and mixed forest biome ALOS-PALSAR HV-polarised backscatter was the most important variable, over temperate conifer forests, Landsat Tasselled Cap variables showed a larger contribution. Yet, in all cases, incorporation of multispectral data improved the retrieval of forest canopy height and reduced the estimation uncertainty for tall forest. An important contribution of this research was the assessment of model transferability for the extrapolation of forest structure at regional or continental scales. The results show that models trained at one study site had higher uncertainty when applied to other sites; however, it was possible to develop a biome model from multiple sites which had similar performance to site-specific models for forest canopy height prediction. This result suggests that a biome-level model developed from several study sites can be used as a reliable estimator of biome-level forest structure from existing satellite imagery.
Finally, our research on the extrapolation of LiDAR-based forest canopy structure also included the assessment of integrating multi-baseline polarimetric interferometric SAR (PolInSAR) information with LiDAR measurements using a machine learning approach.
Quantification of the uncertainty of the forest structure estimates was another objective of ForeStMap. Therefore, we provided a formal uncertainty analysis and error propagation for calculating the uncertainty in our two-step methodology. Thus, we estimated the uncertainty for each pixel using a bootstrapping approach and the 95% confidence interval were reported. We also provided and estimate of the uncertainty by integrating the pixel level errors over the regions of interest and accounting for the spatial autocorrelation of the errors.
ForeStMap has provided a significant progress towards the development of accurate global products of forest structure and its dynamics, for better understanding the carbon cycle, to design forest sustainable management strategies and mitigate greenhouse gas emissions from deforestation and forest degradation, as well as to halt the loss of biodiversity and to reduce fire risk.
With regards the impact of the project, the developed methodology could be readily integrated into national programmes such as LANDFIRE, providing critical information for the development of adequate strategies for fuel management aimed at reducing wildfire occurrence and impacts. Our analyses on the impact of point density and data model used on the LiDAR estimates of forest structure and its dynamics are very important for biomass monitoring and for an effective implementation of climate change mitigation policies such as REDD+ due to its implications for data acquisition costs. The results achieved by ForeStMap are critical for upcoming satellite LiDAR and SAR missions including the ICESat-2/ATLAS (http://icesat.gsfc.nasa.gov/icesat2/) ; the Global Ecosystem Dynamics Investigation (DEGI; https://science.nasa.gov/missions/gedi) ; BIOMASS (https://earth.esa.int/web/guest/missions/esa-future-missions/biomass) ; the NASA-ISRO Synthetic Aperture Radar (NISAR – http://nisar.jpl.nasa.gov/) ; the HyspIRI (https://hyspiri.jpl.nasa.gov/) and EnMAP (http://www.enmap.org/) missions.
More information on the project can be obtained at the project´s website (https://www2.le.ac.uk/departments/geography/research/projects/forestmap) and contacting Dr. Mariano García (Mariano.garcia[at]uah.es; Mariano.gar.alo[at]gmail.com).
The project has been carried out in two phases. The outgoing phase (01/01/2015-31/12/2016), was carried out at the NASA Jet Propulsion Laboratory, California Institute of Technology; and the returning phase (01/01/2017-31/12/2017) was carried out at the Centre for Landscape and Climate Research, University of Leicester.
The first step of the proposed methodology consisted on deriving accurate information of forest structure (aboveground biomass / AGB, canopy height, fractional cover and canopy bulk density) using airborne LiDAR data. Sensor characteristics and surveying parameters affect LiDAR measurements of the spatial distribution of the canopy components, which will impact vegetation structure metrics derived from them. Moreover, LiDAR measurements can be represented using two different data models, the echo-based or the raster model. The data model used along with the surveying characteristics will have an impact on the biophysical properties derived from airborne LiDAR data; therefore, we evaluated the ability of the metrics derived from the two data models used to represent the LiDAR information, as well as the impact of point density (number of laser hits per square meter) on the AGB estimation in three different biomes (temperate, moist tropical and Atlantic forest). In addition to this analysis, the use of multitemporal LiDAR data for forest and carbon monitoring was investigated, including tree growth and biomass / carbon dynamics of a temperate forest site.
The second step of the methodology aimed at extrapolating local airborne LiDAR estimates of forest structure to larger regions based on satellite data, both multispectral and Synthetic Aperture Radar (SAR) imagery, using a machine learning approach. First, the core of the methodology was developed to provide biomass changes caused by a megafire. Landsat-8 OLI data were used to extrapolate LiDAR-based AGB estimates over the burned area to obtain pre-fire AGB using a Support Vector Machine (SVM) regression approach. After subtracting post-fire AGB estimates it was possible to compute the amount of biomass consumed by the fire and the carbon release. This methodology was also applied to derive canopy fuel properties that are critical for fire behaviour. We also evaluated the relation between the derived properties and the burn severity.
The extrapolation of LiDAR-based forest structure estimates was further investigated using multispectral and SAR data. Thus, using LiDAR samples of canopy height over different study areas distributed across two biomes, we developed SVM regression models to extrapolate LiDAR-derived canopy height using Landsat-8 OLI and ALOS PALSAR imagery. It was found that the relative importance of the variables derived from sensors in the model performance varied depending on forest structure and composition. Thus, whereas over the temperate broadleaf and mixed forest biome ALOS-PALSAR HV-polarised backscatter was the most important variable, over temperate conifer forests, Landsat Tasselled Cap variables showed a larger contribution. Yet, in all cases, incorporation of multispectral data improved the retrieval of forest canopy height and reduced the estimation uncertainty for tall forest. An important contribution of this research was the assessment of model transferability for the extrapolation of forest structure at regional or continental scales. The results show that models trained at one study site had higher uncertainty when applied to other sites; however, it was possible to develop a biome model from multiple sites which had similar performance to site-specific models for forest canopy height prediction. This result suggests that a biome-level model developed from several study sites can be used as a reliable estimator of biome-level forest structure from existing satellite imagery.
Finally, our research on the extrapolation of LiDAR-based forest canopy structure also included the assessment of integrating multi-baseline polarimetric interferometric SAR (PolInSAR) information with LiDAR measurements using a machine learning approach.
Quantification of the uncertainty of the forest structure estimates was another objective of ForeStMap. Therefore, we provided a formal uncertainty analysis and error propagation for calculating the uncertainty in our two-step methodology. Thus, we estimated the uncertainty for each pixel using a bootstrapping approach and the 95% confidence interval were reported. We also provided and estimate of the uncertainty by integrating the pixel level errors over the regions of interest and accounting for the spatial autocorrelation of the errors.
ForeStMap has provided a significant progress towards the development of accurate global products of forest structure and its dynamics, for better understanding the carbon cycle, to design forest sustainable management strategies and mitigate greenhouse gas emissions from deforestation and forest degradation, as well as to halt the loss of biodiversity and to reduce fire risk.
With regards the impact of the project, the developed methodology could be readily integrated into national programmes such as LANDFIRE, providing critical information for the development of adequate strategies for fuel management aimed at reducing wildfire occurrence and impacts. Our analyses on the impact of point density and data model used on the LiDAR estimates of forest structure and its dynamics are very important for biomass monitoring and for an effective implementation of climate change mitigation policies such as REDD+ due to its implications for data acquisition costs. The results achieved by ForeStMap are critical for upcoming satellite LiDAR and SAR missions including the ICESat-2/ATLAS (http://icesat.gsfc.nasa.gov/icesat2/) ; the Global Ecosystem Dynamics Investigation (DEGI; https://science.nasa.gov/missions/gedi) ; BIOMASS (https://earth.esa.int/web/guest/missions/esa-future-missions/biomass) ; the NASA-ISRO Synthetic Aperture Radar (NISAR – http://nisar.jpl.nasa.gov/) ; the HyspIRI (https://hyspiri.jpl.nasa.gov/) and EnMAP (http://www.enmap.org/) missions.
More information on the project can be obtained at the project´s website (https://www2.le.ac.uk/departments/geography/research/projects/forestmap) and contacting Dr. Mariano García (Mariano.garcia[at]uah.es; Mariano.gar.alo[at]gmail.com).