Forest Biomass

The ESA GlobBiomass project aims to reduce uncertainties in current estimates of above-ground biomass (AGB) of forests by developing an innovative synergistic mapping approach at five regional sites for the epochs 2005, 2010 and 2015 and for one global map for the year 2010.

Biomass stock and change maps with spatial resolution of 50 – 150 m and with a multi-temporal approach comprising three epochs: 2000 or 2005, 2010 (reference year), and 2015 will be developed for these regions. The regional maps will aim for an overall accuracy of 80% (20% error) or better.

The five regions cover the most important forest biomes:

• Poland: Temperate forest

• Sweden: Boreal forest

• Indonesia: Tropical forest

• Mexico: Tropical forest-woodland transition

• South Africa: Forest-savanna mosaic

We report on progress in the GlobBiomass Task 5 (Regional Biomass Estimation) and describe the regional methods for generating the regional biomass maps and associated uncertainties.

The method to be used for regional biomass estimation in Poland will be based on Random Forest (Breiman, 2001). This machine-learning algorithm combines decision tree predictors such that each tree depends on the values of a random vector sampled independently and with the same distribution for all regression trees in the model.

In Sweden, two main approaches are foreseen: the k Nearest Neighbour (kNN) and the Water Cloud Model (WCM). The kNN algorithm is a non-parametric method used for regression, while the WCM algorithm describes the relationship between forest backscatter and forest variables.

For the Indonesian forest region, regional AGB maps will be derived based on regression modelling. The reference data are extrapolated by relating field observation to airborne LiDAR and the extrapolated maps will be used in the regional biomass mapping from space.

In the Mexican study region, a Maximum Entropy algorithm (MaxEnt) will be used to map AGB. The algorithm is widely used for estimation of species presence, and has recently been used for classification of remote sensing data.

In South Africa, CESBIO’s methodology to map AGB using L-band SAR data and the MIPERS (Multistatic Interferometric Polarimetric ElectroMagnetic) model will be applied. The model simulations are used to study the relationships between the L-band radar backscatter and the forest AGB. The retrieval method has been developed at country scale for Cameroon’s savanna and is being extended to the whole African savannas.

An analysis of accuracy and uncertainty is an inherent part of the work in all study regions. Despite the methodological diversity for the different AGB retrievals our analysis emphasizes the relevance of a spatially discrete description of uncertainty, identifies advantages and delineates cautionary notes on its use.

Forests worldwide are characterized by complex and spatially heterogeneous structures. Tree size distributions are an important outcome of classical forest inventories and have long been used for describing such forest structures and for deriving attributes like forest biomass. There is a high interest to determine key forest attributes from remote sensing. However, detecting tree size distributions has not been reached so far. Here, we present a physical approach for reproducing forest size structures based on stochastic geometry by interpreting forests as tree crown packing systems. The forest packing model is tested for reproducing observed forest structures of two large tropical forest inventories on Barro Colorado Island in Panama and Sinharaja in Sri Lanka (with a total area of 75 ha). The observed tree size distributions can be reproduced successfully as a result of a few simple principles, i.e. site-specific tree allometries, tree mortality, random placement of trees and competition for space. The heterogeneous local structure of forests is also reproduced successfully using our forest packing model. We quantify that most trees grow up to a height of 30 to 50m contributing to the densest height layer of up to 60% filled space. Our approach is an important step towards identifying a minimal set of processes responsible for generating the horizontal and vertical spatial structure of natural forests. Knowledge, especially on the heterogeneous structure of forests derived from our study can potentially enhance the interpretation of remote sensing measurements.

Vast and disastrous forest and peat fires were affecting Indonesia during the dry season in 2015 putting the country on track to be one of the world’s largest carbon emitters. The region was experiencing a very strong El Niño climate phenomenon, similar or even stronger compared with the severest one in 1997/1998. The burned area assessment by multispectral imagery is not only hampered by haze but also by frequent cloud cover during the wet season, which directly starts after the fire season. SAR (synthetic aperture radar) systems are daylight and weather independent as the signal can penetrate through clouds, haze and smoke and are therefore very suitable for delineating burned areas in the tropics. ESA’s Sentinel-1 satellite continuously acquires C-band data, which were used to assess the burned areas in the Indonesian part of Borneo (Kalimantan) by analysing images before and after the fire season. As changes in vegetation cover strongly affect C-band SAR backscatter, changes in pre- and post-fire images were thoroughly investigated using an automatic changed based classification approach. The results clearly showed burned areas due to a decrease in backscatter after the removal of vegetation cover by the fire.

In a next step, the delineated burned areas were intersected with a wall-to-wall SAR derived aboveground biomass (AGB) map from 2010 that was created within the ESA DUE GlobBiomass. The AGB map has a spatial resolution of 100 m and was developed based on LiDAR AGB estimation and ALSO PALSAR mosaic imagery using a stepwise regression modelling approach with backward stepwise selection. Hence, fire emission in Kalimantan for the 2015 El Nino period could be estimated.

An increasing number of regional biomass maps are becoming available in the recent years but their validation remain problematic due to their large area coverage and large mapping unit, while ground observations are available only for limited small sample areas. The accuracy assessment of biomass maps is therefore a crucial but still problematic step and currently there is not a standard validation strategy for biomass products or ground databases designed for this purpose. We present the results of the EU GEOCARBON project on pantropical biomass mapping, the strategy of the ESA GlobBiomass project for characterizing the accuracy of large-scale biomass maps, and introduce a new global reference in situ database: the Forest Observation System (http://www.forest-observation-system.net). Following experiences developed under the CEOS Working Group on Calibration and Validation, the GlobBiomass validation concept consists of an internal uncertainty analysis, an independent validation, a product inter-comparison and a user assessment. We present a methodology to use different types of existing biomass reference data in a harmonized manner and propose the quality criteria to select biomass reference data, the procedures to harmonize and upscale the reference data, and the metrics used to assess product quality. Such metrics consists on product uncertainty (by quantifying the error propagation in the input data), accuracy (by comparing reference data with the biomass products), consistency (by comparing multi-epoch products), and agreement (by inter-comparison with other existing datasets). The accuracy assessment needs to be performed in close cooperation with the user communities and the international scientific community in order to build international confidence in the GlobBiomass products and support their adoption by the user community, and eventually establish a precursor for permanent and operational biomass validation activities. Within this context, the Forest Observation System (FOS) is an international cooperation launched under the auspices of the European Space Agency to establish and maintain a global in-situ forest biomass database and provide ground biomass data measurements for the Earth Explorer mission BIOMASS. This project aims to establish a data sharing policy and define the technical requirements to ensure data comparability. The IFBN will include national and international consortia (e.g. AfriTRON, RAINFOR, CTFS-ForestGEO) that have been devoted to coordinating long-term research with permanent sampling forest plots using robust protocols for accurate tree mapping and measurement, and established databases with a special emphasis on data quality control at the tree level and valorisation of historical databases. The database will include plots with size of 1 ha (preferably) or at least 0.25 ha for which there are detailed and accurate information on plot location, tree density, forest type, tree species composition, wood density, canopy height, and above ground biomass of trees over 10 cm in diameter.

In the recent years, it was recognized that Africa is a major source of interannual variability in global atmospheric CO2, although Africa contributes less than 4 % of the global fossil fuel emissions. In a review of the most recent estimates of the net long-term carbon balance of African ecosystems [1], the uncertainties associated with these estimates are rather high, as shown by the different results indicating that the carbon balance of African ecosystems varies from a sink of about 3.2 Pg.ha-1 to a small source. In Africa, savannas cover roughly 50 % of the continent and represent 46% of the carbon storage, a situation that contrasts greatly with Latin America and Southeast Asia, where savannas represent 16% of the carbon storage. While African wet forests have been relatively spared from massive deforestation events, savanna woodlands, in contrast, were found to be struck by a rapid deforestation, with, for example, a rate of 2.75% reported for Uganda for the 2005-2010 period [2]. The studies mentioned in [1] therefore pointed out the need to consider the contribution of savannas in the overall carbon budget when determining the direction of the African carbon budget (sink or source). So far, above-ground biomass maps have been produced at global or pan-tropical scales using lidar data from GLAS onboard ICESat and auxiliary data such as MODIS [3] [4], at resolutions of 500 m to 1 km. However, these maps focus primarily on regions with high above-ground biomass, and their mutual agreement is very poor in the low biomass areas such as savannas.

Many studies have shown that long wavelength radar data are sensitive to forest biomass, and mapping of above-ground biomass (AGB) has been demonstrated using P- and L-band backscatter data at the HV polarization. Whereas P-band data are expected to be used for a larger range of AGB values, only spaceborne L-band data from ALOS PALSAR are currently available for studies. As a result, L-band has been extensively assessed for estimating AGB. Literature results suggest that there is an AGB saturation level (usually in the range of 75-150 Mg.ha-1) above which the L-band backscatter intensity does not increase with an increase in AGB. As a result, PALSAR data cannot be used to derive AGB of dense tropical forest (up to 300-400 Mg.ha-1), but appear well adapted to the retrieval of AGB of savanna woodlands (typically less than 100 Mg.ha-1).

In this study, we use a mosaic of ALOS PALSAR data to map the above ground biomass (AGB) in savanna ecosystems in Sub-Saharan Africa. The mosaic was produced by the Japanese Aerospace Exploration Agency [5] with data acquired within a single year (2010, mostly between June and August), and contains dual-polarization data (HH and HV). The method to transform the PALSAR mosaic into an AGB map is based on the Water Cloud Model, which relates the backscatter to AGB through a set of 3 parameters. The parameters are estimated by fitting the model with a collection of above-ground biomass estimates obtained from various sources and various countries (Cameroon, Uganda, Mozambique, Botswana and South Africa). A drastic plot selection is required in order to minimize the errors related to the size of the plots and the environmental conditions (topography, soil moisture, etc.) The resulting map improves the current biomass maps available in Africa, both in terms of spatial resolution and of accuracy in the low biomass range.

References

 [1] P. Ciais, A. Bombelli, M. Williams, S. Piao, J. Chave, C. Ryan, M. Henry, P. Brender, R. Valentini, “The carbon balance of Africa: synthesis of recent research studies”, Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences, 369, 2038–2057, 2011.

[2] FAO, “ Global forest resources assessment. Technical Report,” Food and Agriculture Association of the United Nations, 2010

[3] S. Saatchi, N. L. Harris, S. Brown, M. Lefsky, E. T. A. Mitchard, W. Salas, B. R. Zutta, W. Buermann, S. L. Lewis, S. Hagen, S. Petrova, L. White, M. Silman, and A. Morel, "Benchmark map of forest carbon stocks in tropical regions across three continents," Proceedings of the National Academy of Sciences, vol. 108, pp. 9899-9904, 2011.

[4] A. Baccini, S. J. Goetz, W. S. Walker, N. T. Laporte, M. Sun, D. Sulla-Menashe, J. Hackler, P. S. A. Beck, R. Dubayah, M. A. Friedl, S. Samanta, and R. A. Houghton, "Estimated carbon dioxide emissions from tropical deforestation improved by carbon-density maps," Nature Climate Change, vol. 2, pp. 182-185, March 2012.

[5] http://www.eorc.jaxa.jp/ALOS/en/kc_mosaic/kc_map_50. htm