Estimating tropical forest biomass with a combination of SAR image texture and Landsat TM data: An assessment of predictions between regions

https://doi.org/10.1016/j.isprsjprs.2012.03.011Get rights and content

Abstract

Quantifying the above ground biomass of tropical forests is critical for understanding the dynamics of carbon fluxes between terrestrial ecosystems and the atmosphere, as well as monitoring ecosystem responses to environmental change. Remote sensing remains an attractive tool for estimating tropical forest biomass but relationships and methods used at one site have not always proved applicable to other locations. This lack of a widely applicable general relationship limits the operational use of remote sensing as a method for biomass estimation, particularly in high biomass ecosystems. Here, multispectral Landsat TM and JERS-1 SAR data were used together to estimate tropical forest biomass at three separate geographical locations: Brazil, Malaysia and Thailand. Texture measures were derived from the JERS-1 SAR data using both wavelet analysis and Grey Level Co-occurrence Matrix methods, and coupled with multispectral data to provide inputs to artificial neural networks that were trained under four different training scenarios and validated using biomass measured from 144 field plots. When trained and tested with data collected from the same location, the addition of SAR texture to multispectral data showed strong correlations with above ground biomass (r = 0.79, 0.79 and 0.84 for Thailand, Malaysia and Brazil respectively). Also, when networks were trained and tested with data from all three sites, the strength of correlation (r = 0.55) was stronger than previously reported results from the same sites that used multispectral data only. Uncertainty in estimating AGB from different allometric equations was also tested but found to have little effect on the strength of the relationships observed. The results suggest that the inclusion of SAR texture with multispectral data can go someway towards providing relationships that are transferable across time and space, but that further work is required if satellite remote sensing is to provide robust and reliable methodologies for initiatives such as Reducing Emissions from Deforestation and Degradation (REDD+).

Introduction

The amount of carbon stored and sequestered by tropical forests represents one of the greatest uncertainties in understanding their role in the global carbon cycle (Houghton et al., 2000, Malhi, 2010). Quantifying this uncertainty demands methods that can accurately and precisely measure forest carbon dynamics (Brown, 2002), as well as map the geographic extent of forest cover and change over time. Previous attempts at carbon accounting have often been based upon estimating gross emissions, usually taking the form of mapping forest loss with little regard to replacement biomass and thus often overestimating the impact of avoided deforestation on carbon emissions to the atmosphere (UN-REDD, 2008). A more realistic assessment of carbon emissions to the atmosphere would be to estimate net emissions (i.e. carbon emission from deforestation and accumulation of carbon stocks from subsequent vegetation growth). A key driver for this is the United Nations Reducing Emissions from Deforestation and Degradation (UN-REDD+) programme, which explicitly states the need for the development of robust and replicable methods for net accounting of carbon emissions across large areas. Typically, governments and environmental scientists have relied upon official forest statistics to produce global assessments of forests but these suffer from poor temporal coverage, and variable definitions of forest degradation (Grainger, 2010). Estimating above ground forest biomass directly from satellite remote sensing is an attractive tool for deriving net carbon emissions estimates in a systematic and timely fashion, but its potential is yet to be realised operationally. One major problem with the remote sensing of biomass is that of generalizing or transferring knowledge and methods derived from remotely sensed data over time and space (Wilkinson, 1997, Nagendra, 2001, Woodcock et al., 2001).

Numerous studies have made use of remotely sensed data to study forested environments (Franklin et al., 2001, Boyd and Danson, 2005), but such data have not always been able to provide the specific environmental information required by the research and user communities. This is especially so in tropical environments where many have investigated the use of remotely sensed data to estimate tropical forest biomass but with varying degrees of success (e.g. Steininger, 2000, Foody et al., 2001, Foody et al., 2003, Castro et al., 2003, Wang and Qi., 2008, Hill et al., 2011). Even those attempts that have been successful have used methods that may not generalise accurately in time and space. For example, a relationship derived from the accurate prediction of biomass at one site or time may not yield accurate predictions when applied to images of another site and/or acquired at another time with either the same or a different sensor. This is a manifestation of the ‘one time one place’ approach identified by Woodcock (2002) as a common cause of uncertainty in the application of remotely sensed data to estimate forest biophysical properties. The elements of this problem are typically spatial (e.g. generalisation within an image or between imagery of different locations) and temporal (e.g. generalisation between images of one location acquired over a period of time) (Woodcock et al., 2001) and arise typically as a function of concerns with the remote sensing (e.g. consistent radiometric calibration) or ground conditions (e.g. spatial variation in forest structure). Such factors have been shown to significantly affect methods commonly used in previous studies, such as vegetation indices, which are highly sensitive to variation in topography, view angle and atmospheric conditions for example, and whose ability to predict above-ground biomass (AGB) of different tropical forest environments has been found wanting (Foody et al., 2001, Foody et al., 2003). Clearly, if remote sensing is to be a repeatable and consistent source of environmental information at regional scales then such issues must be addressed.

To address the problem of transferring a relationship geographically and develop a more consistently applicable approach to estimating tropical forest AGB with remotely sensed data, Foody et al., 2001, Foody et al., 2003 used an artificial neural network (ANN). This was trained with all six non-thermal Landsat TM wavebands and AGB measured in the field at three different tropical forest locations (one in S. America and two in S.E. Asia). The trained network was then tested to determine if a single ANN model could accurately estimate biomass at each site. When trained with data from a single site, the network was unable to accurately estimate AGB at the ‘unseen’ sites and performed only marginally better than a number of vegetation indices. More accurate predictions were obtained when the network was trained with data from all three sites (r = 0.38, significant at the 95% confidence level). The strongest correlation (r = 0.49, significant at the 99% confidence level) was observed when the network was trained with training data from all three sites and an additional variable identifying from which site the individual training samples were taken. Thus, a network trained with pixel values from six Landsat TM wavebands, corresponding AGB value and a label indicating the location of each sample, yielded a moderately strong correlation between estimated and observed AGB from all three sites.

Whilst such an approach appears promising, it is possible that an additional independent source of data that is able to discriminate between different forest locations by taking into account differing forest structure and other biophysical differences, could further improve the ability to estimate AGB at many sites concurrently and enhance spatial transferability (i.e. the transferring of predictive relationships between sites). Such independent data need to be widely available and should provide information in addition to that contained within multispectral data. This information could potentially be derived from active remote sensing systems which have been shown to be highly correlated with AGB in some environments (Brown, 2002, Lu, 2006, Mitchard et al., 2009).

LiDAR systems show particular promise in this respect (e.g. Hyde et al., 2006), but the availability of systematically acquired high-density LiDAR data for tropical regions remains very limited. Perhaps more useful at the present time is synthetic aperture radar (SAR). SAR data have been successfully used to estimate AGB in many forested environments, both by using the data to classify different forest types or by direct estimation. When classification-based methods have been used it has often been shown that texture information derived from SAR images can be especially useful in discriminating between different forest classes (e.g. Podest and Saatchi, 2002) as image texture contains information on the structural and geometric properties of forest canopies (DeGrandi et al., 2009). Direct estimation of AGB has been largely based upon the fact that SAR backscatter (σ0) is sometimes strongly correlated with forest biomass, particularly in low-medium biomass forests and at lower frequencies (P- and L-band) (e.g. Le Toan et al., 1992, Luckman et al., 1997, Castel et al., 2002, Lucas et al., 2006a). Again, the addition of image texture measures has been shown to improve the accuracy with which biomass can be estimated in regenerating tropical forests (Kuplich et al., 2005). However, the use of SAR data to directly estimate forest AGB in tropical regions has well-known limitations, especially the problem of backscatter saturation at relatively low AGB. Previous studies have reported saturation at AGB of around 20 tha−1 for C-band and 40 tha−1 for L-band (Imhoff, 1995), although these may be extended through the use of backscatter ratios (Foody et al., 1997). The saturation effect is compounded by uncertainty in the relationships between AGB and SAR data resulting largely from a reliance on plot-based estimates of AGB that often fail to take into account variability in radar parameters, topography and forest structure (Luckman et al., 1997, Lucas et al., 2006a, Lucas et al., 2006b). It is likely that such issues have restricted investigation into the use of SAR data to estimate tropical biomass previously.

Several studies have investigated the integration of SAR and multispectral remotely sensed data for the estimation of forest biophysical properties. These have tended to employ a combination of optical and SAR data to aid discrimination of different vegetation and forest classes, from which typical AGB values for each class can be aggregated with respect to the entire classified area (e.g. Amini and Sumantyo, 2009). Another form of integration is to use SAR data as an extra variable (alongside multispectral wavebands and/or vegetation indices) in multivariate regression models. For example, Rauste (2005) used a combination of JERS-1 SAR backscatter and Landsat TM data to directly estimate stem volume of forests in Finland. This resulted in a slight increase in the accuracy of stem volume estimation compared to when SAR data alone were used (r increased from 0.85 to 0.89). As complementary data to multispectral Landsat TM imagery, therefore, the synergistic use of SAR and multispectral data for estimating AGB appears promising and deserves further attention (Lu, 2006).

Testing the spatial transferability of predictive relationships between remotely sensed data and AGB has received little attention to date. Mitchard et al. (2009) showed that a relationship derived between AGB and L-band backscatter was successful in estimating AGB in four different savannah and low biomass tropical forest environments in Africa. In high biomass tropical forests Foody et al. (2003) demonstrated that whilst AGB could be estimated at single sites reasonably well, poor accuracies were observed when relationships were used to estimate AGB at other sites. Whilst both the above studies used single source remotely sensed data, the aim of this paper is to investigate whether a combination of widely available multispectral (Landsat TM) and L-band SAR (JERS-1) backscatter and image texture measures can be used to improve the spatial transferability of predictive relationships at three high-biomass tropical forest sites, as opposed to using multispectral data on their own.

Section snippets

Test sites and data

To test whether a combination of SAR and multispectral data could estimate AGB at different tropical forest locations empirical relationships were derived between ground estimates of AGB and the remotely sensed data. Forest plots were located at three different test sites, located near Manaus in Brazil (2° 28′ S, 59° 58′ W, Danum Valley Field Centre in Borneo, Malaysia, (4° 50′ N, 117° 45′ E) and part of the Khun Khong catchment in north-west Thailand (19° 31′ N, 98° 48′ E). All sites are

Data processing

All remotely sensed data for each site were processed in the same way, before then being used to estimate AGB.

Results

Multiple individual ANNs were created in each of the four training scenarios, with each individual network varying in their network architecture, type and data used as inputs (although learning and momentum rates were kept constant throughout). Results are reported in all cases from the network that showed the strongest correlation with the testing data.

SAR texture

The two independently derived measures of SAR image texture (wavelet and GLCM-based) were both highly correlated with AGB when neural networks were trained and tested with texture samples from the same sites only. In fact, the strength of the correlations observed are comparable to those reported previously (Foody et al., 2001, Foody et al., 2003), when neural networks were trained and tested with Landsat TM data only from the same sites (r = 0.82, 0.71 and 0.84 for Malaysia, Thailand and Brazil

Summary and conclusions

A key requirement in the construction of global knowledge in relation to tropical forests is that methods and relationships should be repeatable and applicable across different tropical forest types and locations. A common argument for the use of remote sensing to generate this knowledge is that it should be able to provide estimates of forest properties in a consistent manner, given the systematic way in which data are collected and potential for standard processing routines to be used. This

Acknowledgements

The field and optical remotely sensed data for the Malaysian and Thailand test sites were acquired for the EU-funded INDFORSUS project (ER-BIC18T960102), while the field and optical data for the Brazilian test site were acquired through the NERC TIGER project (GST/02/604). We are grateful to all involved with these projects for their assistance and input. The acquisition and processing of the JERS-1 data was made possible as a result of financial assistance from the Carnegie Trust for the

References (56)

  • Y. Rauste

    Multi-temporal JERS SAR data in boreal forest mapping

    Remote Sensing of Environment

    (2005)
  • G. Sun et al.

    Radiometric slope correction for forest biomass estimation from SAR data in the Western Sayani Mountains, Siberia

    Remote Sensing of Environment

    (2002)
  • C.E. Woodcock et al.

    Monitoring large areas for forest change using Landsat: generalisation across space, time and Landsat sensors

    Remote Sensing of Environment

    (2001)
  • J. Amini et al.

    Employing a method on SAR and optical images for forest biomass estimation

    IEEE Transactions on Geoscience and Remote Sensing

    (2009)
  • R. Baishya et al.

    Estimation of tree biomass, carbon pool and net primary production of an old growth Pinus Kesiya Royle ex. Gordon forest in North-eastern India

    Annals of Forest Science

    (2011)
  • T.R. Baker et al.

    Increasing biomass in Amazonian forest plots

    Philosophical Transactions of the Royal Society of London Series B

    (2004)
  • R.J. Bijlsma

    The characterisation of natural vegetation using first-order and texture measurements in digitized, color-infrared photography

    International Journal of Remote Sensing

    (1993)
  • C.M. Bishop

    Neural networks for pattern recognition

    (1995)
  • D.S. Boyd et al.

    Satellite remote sensing of forest resources: three decades of research development

    Progress in Physical Geography

    (2005)
  • Brown, S., 1997. Estimating biomass and biomass change of tropical forests. FAO Forestry Paper, vol. 134. Rome,...
  • S. Brown et al.

    Biomass estimation methods for tropical forest with application to forest inventory data

    Forest Science

    (1989)
  • K.L. Castro et al.

    Monitoring secondary tropical forests using space-borne data: implications for Central America

    International Journal of Remote Sensing

    (2003)
  • J. Chave et al.

    Tree allometry and improved estimation of carbon stocks and balance in tropical forests

    Oecologia

    (2005)
  • P.S. Chavez

    Radiometric calibration of Landsat Thematic Mapper multispectral images

    Photogrammetric Engineering and Remote Sensing

    (1989)
  • P.S. Chavez

    Image-based atmospheric corrections: revised and improved

    Photogrammetric Engineering and Remote Sensing

    (1996)
  • S. Ekstrand

    Landsat TM-based forest damage assessment: correction for topographic effects

    Photogrammetric Engineering and Remote Sensing

    (1996)
  • Daubechies, I., 1991. The wavelet transform: a method for time–frequency localization. In: Haykin S. (Ed.), Advances in...
  • G. DeGrandi et al.

    Analysis by wavelet frames of spatial statistics in SAR data for characterising structural properties of forests

    IEEE Transactions on Geoscience and Remote Sensing

    (2009)
  • Cited by (179)

    View all citing articles on Scopus
    View full text