Upscaling Forest Biomass from Field to Satellite Measurements: Sources of Errors and Ways to Reduce Them
- 312 Downloads
Forest biomass monitoring is at the core of the research agenda due to the critical importance of forest dynamics in the carbon cycle. However, forest biomass is never directly measured; thus, upscaling it from trees to stand or larger scales (e.g., countries, regions) relies on a series of statistical models that may propagate large errors. Here, we review the main steps usually adopted in forest aboveground biomass mapping, highlighting the major challenges and perspectives. We show that there is room for improvement along the scaling-up chain from field data collection to satellite-based large-scale mapping, which should lead to the adoption of effective practices to better control the propagation of errors. We specifically illustrate how the increasing use of emerging technologies to collect massive amounts of high-quality data may significantly improve the accuracy of forest carbon maps. Furthermore, we discuss how sources of spatially structured biases that directly propagate into remote sensing models need to be better identified and accounted for when extrapolating forest carbon estimates, e.g., through a stratification design. We finally discuss the increasing realism of 3D simulated stands, which, through radiative transfer modelling, may contribute to a better understanding of remote sensing signals and open avenues for the direct calibration of large-scale products, thereby circumventing several current difficulties.
KeywordsBiomass Calibration Carbon Error propagation Field data Modelling
The large uncertainty associated with the global spatio-temporal dynamics of forest carbon (C) is a major obstacle to the projection of future atmospheric CO2 concentrations and the implementation of mitigation strategies. Satellite observations of land-use changes are currently reliable enough to provide robust information on deforestation dynamics (Hansen et al. 2013). However, the associated C release remains uncertain because broad-scale C mapping applications from satellite data still convey large uncertainty (Mitchard et al. 2013; Huang et al. 2015; Rodriguez-Veiga et al. 2017). Even higher uncertainty is associated with C fluxes related to forest degradation and regrowth because small changes in the C stocks in closed-canopy forests are challenging to detect remotely (Bustamante et al. 2016), resulting in an underestimation of forest degradation globally (Pearson et al. 2017).
2 Improving Field-Based AGB Estimation
Uncertainty in field-based AGB estimation has recently become a matter of concern for the RS community (e.g., Mermoz et al. 2014, 2015; Longo et al. 2016; Xu et al. 2017; Bouvet et al. 2018; Jucker et al. 2018a). For instance, Chen et al. (2015) developed an analytical framework to track the sources of errors from field measurements in airborne LiDAR-AGB predictions. The authors estimated that field-based uncertainty contributed only 10% of the total pixel-level uncertainty at a 0.16-ha resolution (see also Longo et al. 2016). However, the statistical framework developed by these authors rested upon the assumption that tree-level AGB estimation errors are independent within and between field plots; thus, they average out in RS models. While this assumption is common (e.g., Chave et al. 2004; McRoberts and Westfall 2013; Longo et al. 2016), it is often violated, generating overlooked biases in AGB maps.
Field-based AGB estimates at the plot level are obtained by summing individual AGB values for all trees within the plot. At the tree level, AGB is generally estimated using an allometric model that combines key tree features, typically the stem diameter (D), total tree height (H) and wood density (WD; Chave et al. 2005, 2014). While D may suffer from important measurement errors (Clark 2002), significant efforts have been made towards protocol standardization (Phillips et al. 2009; Condit et al. 2014) to warrant negligible errors, at least in scientific plots (Chave et al. 2004; Molto et al. 2013). However, as shown below, the two other predictors, as well as the AGB model itself, may convey systematic error components within and between plots that may directly propagate into RS models.
2.1 Uncertainty in Wood Density
Individual tree AGB is directly proportional to WD, which varies by as much as tenfold among tree species (Chave et al. 2009). However, WD is rarely measured on individual trees in forest inventories, and in most cases, a mean specific value is assigned from independent databases to all trees of a given taxonomic category, often higher than the species level. This approximation is very crude beyond the genus level (Flores and Coomes 2011); however, it is common in the tropics because of inaccurate botanical identifications (Gomes et al. 2013) or the limited availability of WD data (Chave et al. 2009). For example, in the well-known pantropical forest inventory dataset generated by Al Gentry, which includes 43,000 trees (> 8000 taxa), more than one-third of the trees could not be assigned a WD value at the species or genus level from the Global Wood Density Database (Chave et al. 2009; Zanne et al. 2009). Given that most tropical tree taxa are spatially aggregated at various scales (Condit et al. 2000), uncertainty in WD estimates is expected to cause spatially structured errors, potentially generating biases in plot-level AGB estimates (Baker et al. 2004). One way to mitigate this error source is to rely on permanent scientific sites for model calibration. For instance, in the 50-ha Center for Tropical Forest Science (CTFS) permanent plot at Barro Colorado Island, Panama (Condit 1998), less than 2% of the trees have WD values assigned at a coarser taxonomic resolution than the genus level. In addition, repeated censuses also strongly minimize field measurement or encoding errors and reduce the number of missing trees (Lopez-Gonzalez et al. 2011), providing high-quality calibration data for RS applications (Chave et al. this issue).
The assignment of a WD value to an individual tree at the species level does not, however, ensure bias-free estimates. WD is known to vary widely between individuals of the same species and even within individual trees (Tarelkin et al. 2019; Swenson and Enquist 2008; Bastin et al. 2015a; Wassenberg et al. 2015). For example, species having at least 10 observations in the Global Wood Density Database (n = 109) exhibit a mean WD variation of 9% around the mean. Given that two-thirds of the taxa in that dataset are documented with a single measure, the effects of intra-specific variations in WD on AGB estimates are currently difficult to quantify. Intra-individual variations may also have a strong impact on field AGB estimations, e.g., when converting volume estimations (e.g., from terrestrial LiDAR, see below) into biomass. A recent study conducted in Cameroon showed a significant decrease in WD with height for most species; thus, the use of WD at the trunk base to convert volume into biomass led to an AGB overestimation of approximately 10% at the plot level (Sagang et al. 2018). The authors, however, demonstrated that vertical variations could be predicted from basal WD; thus, these variations could be controlled for in biomass estimations, which was recently confirmed in a large dataset from 6 different countries in central Africa (Momo Takoujdou et al. in preparation).
A lesson learned from this short review is that more efforts are needed to collect WD data for tropical taxa using standardized protocols (Williamson and Wiemann 2010) while accounting for all major sources of variation. Given the time and cost of wood core analyses, non-destructive and rapid WD measurement techniques are appealing, such as torsiometers, Pilodyns, nail withdrawal tools or emerging electronic devices to measure drilling resistance (reviewed in Gao et al. 2017). For instance, the use of an empirical model developed from Pilodyn measurements led to a predicted WD error of 15% for 1427 trees from four continents.
2.2 Uncertainty in Tree Height
Beyond measurement errors, H–D allometric models represent a major source of uncertainty in field-based AGB estimates because this allometry exhibits large spatially structured variations at local, regional and continental scales (Ketterings et al. 2001; Feldpausch et al. 2011; Vieilledent et al. 2012; Vincent et al. 2012b; Réjou-Méchain et al. 2015). To account for broad-scale variations, both regionally averaged H–D relationships (Feldpausch et al. 2012) and bioclimatic proxies (Eq. 6 in Chave et al. 2014) have been proposed. These approaches significantly reduce AGB model errors in tropical regions, but certain local deviations may remain quite large. For instance, in the central Congo Basin, forest AGB was overestimated by 24% when the Feldpausch et al. (2012) regional H–D relationship was used instead of a local relationship (Kearsley et al. 2013). In the Khao Yai Forest Reserve in Thailand, we found that the bioclimatic proxy proposed by Chave et al. (2014) underestimated H by 26% (Jha, Chanthorn, Réjou-Méchain, unpublished). Finally, both approaches resulted in an AGB underestimation > 20% in a Brazilian forest (Hunter et al. 2013). Considering such uncertainty in only H estimations, any RS–AGB model calibrated in these regions would result in a systematic bias > ± 20% in AGB estimations (case B in Fig. 1), i.e. an error margin larger than the commonly admitted accuracy.
H–D relationships may also strongly vary between plots within landscapes. Average differences in H of approximately 10% for a given D were found between plots within areas < 10 km2 in French Guiana (Vincent et al. 2012a, 2014; Réjou-Méchain et al. 2015). More specifically, the H–D relationship was found to strongly vary along a successional gradient in Thailand (Chanthorn et al. 2017); hence, this relationship varied concomitantly with forest AGB, potentially leading to a non-uniform bias in RS models (case of Fig. 1c). Various strategies may be adopted to better account for these inter-plot variations. The most straightforward strategy is to acquire H data for at least a subsample of trees to calibrate plot-specific H–D allometries (Vieilledent et al. 2012; Hunter et al. 2013), even though care must be taken with respect to the H measurement method and sampling strategy (see above and Sullivan et al. 2018). Another way to better account for systematic variations in H–D allometry between plots would consist of finding local covariates. Forest structural metrics that reflect stand competition intensity or soil fertility were found to explain some variation, but to a relatively small extent (Feldpausch et al. 2011; Banin et al. 2012). In contrast, mean canopy height has been found to correlate well with tree slenderness (a higher H for a given D), both within species and at the community level (Vincent et al. 2012a). Hence, when a high-resolution canopy height map is available, the accuracy of individual H estimations from D measurements can be significantly improved by adding local canopy height metrics to the H–D model (e.g., median of the canopy height model in a 50-m window centred on each tree). Using this approach, the individual H error was reduced by 14%, leading to a threefold decrease in the plot-level AGB bias at the Paracou site, French Guiana (n = 2141 trees; Vincent et al. unpublished). This result adequately illustrates that if field-based estimates generally feed into RS models, they may also themselves be improved by RS measurements. With the recent advances in our ability to extract individual trees from dense LiDAR point clouds (Ferraz et al. 2016), H measurements may be directly assigned to individual canopy trees that account for a large share of the stand AGB (Bastin et al. 2015b). In conclusion, better integration of LiDAR technologies and field measurements has the potential to substantially reduce the uncertainties associated with tree height variations.
2.3 Uncertainty in AGB Allometric Equations
The allometric conversion of individual tree attributes into AGB is a major source of uncertainty in tree and plot AGB estimates (Chave et al. 2004; Molto et al. 2013). At the tree level, the predictive error associated with the pantropical allometric model of Chave et al. (2014) reaches almost 40% (the total error is slightly above 50%), but if considered random, this error rapidly averages out as the number of trees increases in the predicted population (Réjou-Méchain et al. 2017). A cross-validation assessment of this model indeed revealed a mean bias across sites of only 5% but exceeded 30% in 8 out of 58 sites (Chave et al. 2014). Models that assume universal allometry may thus be liable to strong and uncontrolled local bias, which is generally ignored in error propagation schemes. This site effect may reflect differences in species composition or biotic (e.g., competition intensity) or abiotic (e.g., soil fertility) conditions. For instance, the use of biome-specific rather than species-specific allometries resulted in an AGB prediction bias between − 18% and + 14% for a Brazilian mangrove forest (de Souza Pereira et al. 2018).
Crown allometry variations among sites have also been identified as a major source of uncertainty in pantropical allometric models (Goodman et al. 2014). The proportion of the crown to the total tree AGB is indeed highly variable among trees (3–88%) and tends to increase with the increase in tree AGB beyond 10 Mg, reaching 50% on average for trees > 45 Mg and 34% for trees < 10 Mg (Ploton et al. 2016). This systematic change in tree shape with total tree AGB explained 20% of the underestimation of the AGB of large trees reported by Chave et al. (2014). Using simulations, Ploton et al. (2016) showed that, for 1-ha plots, this bias led to systematic errors ranging from − 26% to + 16% that depended on the frequency of large trees, thus introducing a non-uniform bias in AGB maps (Fig. 1c). Integrating crown dimensions in allometric models strongly reduces this allometric uncertainty. Jucker et al. (2017b) even proposed an AGB allometry using crown size and tree height, i.e. tree dimensions potentially directly measurable by means of RS (e.g., Ferraz et al. 2016).
2.4 Perspective on Field-Based AGB Estimation
The use of terrestrial laser scanners (TLS) has recently emerged as a credible alternative to destructive approaches for the estimation of tree AGB (Calders et al. 2014; Disney 2018; Gonzalez de Tanago et al. 2018; Lau et al. 2018; Momo Takoudjou et al. 2018). This technology and the associated processing methods are rapidly progressing but still face a number of challenges in tropical moist forests, such as the high degree of occlusion, the difficulty in segmenting individual trees in intricate vegetation (Trochta et al. 2017; Calders et al. 2018) and in filtering out leaves (Béland et al. 2014; Ma et al. 2016; Calders et al. 2018), which is a required step prior to tree structure reconstruction from the point cloud (Momo Takoudjou et al. 2018). Overcoming these issues would take us one step closer to an automated routine for reconstructing entire forest stands, thereby producing large amounts of data to feed allometric models (Raumonen et al. 2013). However, the above-mentioned problems associated with poor WD estimates and the presence of hollow trees remain a serious difficulty for TLS to overcome, which inherently measures the outer volume.
Another perspective from the advent of LiDAR technology is the possibility of directly estimating stand-level AGB instead of summing the AGBs of individual trees. For instance, Vincent et al. (2014) analytically derived a stand-level AGB equation based on stand volume-weighted WD, tree density, mean quadratic diameter and mean canopy height, which, as discussed above, controls the local H–D allometry. Stand density and mean quadratic diameter can both be extracted from TLS data (e.g., Bauwens et al. 2016) more accurately and more easily than individual tree volumes, while a canopy height model within a scanned plot can be derived from the same TLS data. Once such a relationship is calibrated, airborne laser scanners (ALS, including those on unmanned aerial vehicles (UAVs)) suffice to derive AGB estimates from standard stem diameter measurements and WD estimates. However, the calibration step is challenging because it would ideally rely on destructive stand-level measurements. Such an approach would probably limit the propagation of individual errors, especially when neither tree height measurements nor a reliable local H–D allometry is available.
3 Matching Field and RS Measurements
When an RS signal is related to a field-based estimate of forest AGB, the basic assumption is that they both measure the same area and objects. This assumption is often violated in practice, leading to large uncertainty in the final AGB density maps (Gobakken and Naesset 2009; Frazer et al. 2011; Mascaro et al. 2011; Réjou-Méchain et al. 2014; Saatchi et al. 2015). Here, we review important sources of mismatch between field and RS measurements and identify ways to overcome them.
3.1 Geolocation Uncertainty
3.2 Acquisition Angle
Satellite-borne instruments do not always measure forests vertically (i.e. in nadir mode). When an active penetrating signal (e.g., RADAR or LiDAR) is sent with a large (pluri-metric) footprint size and a significant incidence angle (typically > 30° for RADAR sensors; Saatchi et al. 2011a; Robinson et al. 2013), the forest volume measured in field plots and intercepted by the satellite sensor does not correspond to the same physical objects, especially for small plots or in topographically complex areas (Villard and Le Toan 2015). Using theoretical simulations, Réjou-Méchain and Barbier (unpublished) found that an incidence angle of 30° may lead to volume differences of ca. 55, 20 and 10% for plot and pixel sizes of 20, 50 and 100 m, respectively, even if existing surrounding trees partly compensate for this mismatch. Thus, any RS signal acquired with a significant incidence angle should not be calibrated and validated with small field plots or in small patches of heterogeneous vegetation.
3.3 Mismatch in Tree Representation
The mismatch between the RS and field measurements may also be due to differences in tree representation. Conventional field measurements for AGB estimations are “trunk-based”, i.e. a tree is considered to belong to a given field plot when half its trunk base section is within the plot limits, irrespective of the proportion of its vertical crown projection falling outside the plot. With RS approaches, AGB is measured from an area-based perspective, i.e. only the plant material with a ground projection within the area of interest is considered. Using LiDAR data, Mascaro et al. (2011) showed that accounting for this edge effect reduces the LiDAR-AGB model error by 55, 21 and 4% at 20-, 50- and 100-m resolution, respectively, i.e. with a decrease in the perimeter-to-area ratio, the associated edge effects logically vanish because the proportion of trees with crowns crossing the plot edge decreases with the increase in plot size.
Differences in tree representation may also impact the estimates of forest C dynamics. Réjou-Méchain et al. (2015) compared the 4-year AGB dynamics inferred from repeated field plot censuses and LiDAR acquisitions and found relatively poor agreement between the two estimates; these authors concluded that different components of forest dynamics are measured by the two approaches. Natural canopy dynamics is dominated by many small-scale events that are captured by LiDAR data but not by field measurements (Kellner and Asner 2009; Leitold et al. 2018). Similarly, leaf density, which influences LiDAR estimates but not field estimates, regenerates faster than woody biomass after a disturbance (Asner et al. 2006). Thus, any dynamics inferred from RS measurements should be interpreted differently from the dynamics inferred from the ground.
3.4 Temporal Mismatch
The temporal difference between RS and field measurements is a common problem in RS studies. The impact of such temporal mismatches is difficult to predict in natural forests due to the high stochasticity of tree mortality. This stochasticity, however, decreases with the increase in plot size (Chambers et al. 2013); hence, the error associated with the temporal difference between measurements is likely to decrease with increasing plot size in the absence of major climate anomalies (e.g., El Niño events) or major disturbances, such as hurricanes or fires, which may be detected with Landsat (Kennedy et al. 2010) or MODIS (Justice et al. 2002) time series data. (MODIS is the Moderate Resolution Imaging Spectroradiometer aboard the NASA Terra and Aqua satellites.) In this case, estimated growth rates may be used to compensate for the time lag in AGB estimates from field and RS data (Avitabile and Camia 2018).
3.5 Scale Mismatch
The issue of scale mismatch between calibration field plots and RS data pixels is challenging, particularly when coarse-resolution RS products are calibrated with numerous small field plots such as national forest inventory data collected in sampling units less than 0.1 ha in size. For instance, the pantropical 1-ha resolution AGB map of Saatchi et al. (2011b) was calibrated using field plots that were typically ≤ 1 ha in size, hence representing a local sampling rate of 1% at best. For instance, Réjou-Méchain et al. (2014) showed that the use of calibration plots smaller than the RS pixels generates large sampling errors, producing significantly biased AGB maps (case Fig. 1c). For a given pixel-to-plot size ratio, the error due to local AGB variability is larger for small plots and pixels than for large plots and pixels (Réjou-Méchain et al. 2014). Thus, studies aiming at building coarse-resolution maps face the challenge of minimizing this sampling error. A common practice is to spatially aggregate field estimates to better match the RS pixel resolution; however, this approach results in large uncertainty in the associated sampling error (Avitabile and Camia 2018). A more reliable approach consists of using intermediate-resolution RS data, such as very high-resolution images, to bridge the gap.
4 Bridging Field and Satellite-Based Measurements
Field plots are generally too costly and time-consuming to establish to be densely and evenly distributed within landscapes. Consequently, RS models are often calibrated with field plots that are not representative of the area of interest (Marvin et al. 2014). Furthermore, field plot size rarely approaches the size of coarse-resolution RS pixels (e.g., MODIS, ca. 250 m), which are the current basis of the wall-to-wall biome or global mapping. This problem generates the mismatch issues discussed in the previous section. To overcome these problems, a multi-step upscaling approach involving RS data of very high spatial resolution (VHSR) can provide intermediate-scale biophysical maps to be then used as reference to calibrate wall-to-wall RS data of coarse resolution and broad swaths (Asner et al. 2013; Baccini and Asner 2013; Xu et al. 2017). In this section, we first review the VHSR RS products that may be used in such an approach and then discuss the caveats and perspectives associated with such a multi-step upscaling strategy.
4.1 Airborne LiDAR
Airborne LiDAR (ALS), with its ability to penetrate the canopy, provides a fine 3D description of the forest structure and represents an excellent option for linking field plots to broad-scale RS data. Over the last decade, a large number of studies have aimed at retrieving forest AGB from LiDAR-derived metrics, usually with good accuracy at the 1-ha scale (i.e. with a model prediction error typically below 15%; Zolkos et al. 2013). ALS allows for the computation of a variety of metrics related to the vertical or horizontal forest profile (Lefsky et al. 2002). Many studies have relied on statistical fittings (e.g., stepwise linear models) to identify the most predictive metrics in a given context, resulting in site-specific ALS–AGB model forms with limited transferability across sites (Vincent et al. 2012b; Zolkos et al. 2013). For this reason, some studies aimed at designing generic ALS–AGB models that were expected to perform consistently across sites. For instance, Bouvier et al. (2015) identified predefined metrics of strong complementarity and interpretability to build a generic model transferable across temperate forest types. In the tropics, another attempt used a single ALS metric, the mean top-of-canopy height (TCH), combined with minimal field data to predict forest AGB. Asner and Mascaro (2014) used data from 14 tropical areas to calibrate a model of the form AGB = aTCHb1BAb2 WDb3, where the basal area (BA) was locally predicted from TCH, and where WD relied on regional estimates. However, the fitted equation, which was presented as universal, led to underestimations of 7% (Jucker et al. 2017a) and 16% (Réjou-Méchain et al. 2015) in two independent sites compared with locally adjusted models. A modification of the model form was suggested by Vincent et al. (2014), where the scaling of AGB estimates from tree to plot used the stem number (N) and average stem cross-sectional area (or quadratic diameter) instead of BA to avoid unwarranted errors in the scaling process. A universal predictive equation that relates ALS metrics and AGB should continue to be sought to maximize the benefits from the increasing availability of ALS data (Labriere et al. 2018). However, as long as H–D variability is not accounted for, a single predictive equation would remain elusive. Furthermore, airborne data acquisition remains costly (ca. 200–500€ km−2) and is subjected to flight authorization, which hampers ALS acquisitions in certain countries.
4.2 Unmanned Aerial Vehicle (UAV) Systems
4.3 VHSR Spaceborne Systems
Spaceborne data may be mobilized to complement ALS or UAV data and move beyond the relatively small extents of ALS and UAV campaigns, notably because these data can be acquired at a low cost and with few issues regarding authorizations. For instance, VHSR satellite optical images (≤ 2 m resolution) are able to characterize canopy texture, which informs the size distribution of canopy crowns and inter-crown gaps (Couteron et al. 2005; Frazer et al. 2005), thus indirectly informing stand structure and AGB. Texture has historically been widely used in forest science for visual interpretation of aerial photographs. Similar interpretations from automated processing can be carried out from VHSR images via canopy texture analysis whenever images have sufficient effective resolution (the maximum acceptable seems to be approximately 2 m; Proisy et al. 2007). For instance, the Fourier-based textural ordination (FOTO) method aims to ordinate canopy image windows along texture gradients based on Fourier spectra (Couteron et al. 2005), and this method successfully retrieved the AGB gradients in several case studies across the tropics with an accuracy only slightly lower than that of ALS approaches (relative prediction error ≤ 20%; Proisy et al. 2007; Ploton et al. 2012; Bastin et al. 2014; Singh et al. 2014; Pargal et al. 2017) but at 50–100 times lower cost (2–10€ km−2 for optical images). Hence, the trade-off between data interpretability versus affordability should suggest designing scaling-up strategies based on nested sampling of field, airborne and VHSR satellite data. Satellite-borne optical sensors offering VHSR may thus be part of an effective chain for mapping AGB at landscape-regional scales. However, there are also situations in which texture analyses fail to properly retrieve certain AGB gradients (Ploton et al. 2013; Blanchard et al. 2015) because the way in which texture features relate to AGB strongly varies across forest types (Ploton et al. 2017). Due to the limited number of field plots available for calibration and validation, biophysical drivers of canopy texture have not been fully identified in previous studies. The increasing availability of ALS data will help to better evaluate such texture-based approaches and ultimately allow for the calibration and validation of texture indices in sufficiently diverse forest types. However, canopy texture, such as reflectance, is strongly impacted by the Sun–scene–sensor geometry, although correction techniques have been proposed (Barbier et al. 2011; Barbier and Couteron 2015). The number of textural descriptors, therefore, needs to be kept to a minimum to control these effects in large collections of images acquired in diverse geometrical configurations.
Apart from texture indices, stereophotogrammetric approaches from VHSR satellite images allow for the retrieval of canopy surface models that may then be used to predict AGB. This retrieval implies more cloud-free images and a higher budget for data acquisition and, contrary to LiDAR, stereophotogrammetric approaches do not provide information on fine-scale topographical variation, which is detrimental for building canopy height models. Thus, such approaches remain very scarce in the literature (St-Onge et al. 2008; Lagomasino et al. 2015). Finally, non-penetrating X-band RADAR satellite images used in stereo mode (Tandem-X) have also shown to have the potential to predict AGB gradients linked to logging and forest degradation and to assess AGB losses through diachronic comparisons (Schlund et al. 2015; Solberg et al. 2018).
4.4 Limits and Perspectives on Bridging Field and Satellites
Regardless of the VHSR product or model structure used for bridging field plots and large-scale RS data, there are several caveats. First, both the number of calibration field plots and their size remain important for restricting VHSR-based AGB prediction errors to acceptable limits. For instance, in a meta-analysis of LiDAR-based AGB estimations, Zolkos et al. (2013) showed an asymptotic decrease in the prediction error with the increase in the size of the calibration plots, with a slow decay on average above a plot size of 0.2 ha, corresponding to a prediction error of ca. 20%. However, highly contrasted results can be found across case studies, suggesting that prediction errors vary with forest structure at a given plot size. For instance, in tropical forests of Ghana, Chen et al. (2015) found a LiDAR-based prediction error on AGB of ca. 42% at a spatial resolution of 0.16 ha. Thus, AGB estimates derived from VHSR data and used for calibrating broad-scale products should be performed at a rather coarse resolution, with 1 ha being a conservative resolution to ensure an acceptable accuracy.
Another major problem that may arise during the plot to VHSR step is that significant biases may occur during the extrapolation of plot AGB estimates through the VHSR products. For instance, the prediction of AGB from local stand height or related RS metrics relies on knowledge of local WD and height–diameter allometry. As discussed in Sect. 2, both measures are known to vary, sometimes abruptly, within landscapes, and there is thus no guarantee against biases when extrapolating AGB estimates over areas where these variations are unknown. Stratifying the landscape by forest types prior to AGB model calibration thus appears essential to guarantee model robustness, although there are not many accounts of thorough pre-stratification processes in the literature (Gregoire et al. 2016). However, sources of complementary data are developing. Notably, hyperspectral sensors now appear to able be to provide information on canopy species composition and help infer WD (Jucker et al. 2018b). Moreover, some LiDAR metrics (e.g., canopy gap fractions) can help assess the disturbance or successional status of the forest with relevance for the explicit estimation of WD (Guitet et al. 2018).
5 Global Wall-to-Wall Extrapolation
Earth system models and international policy initiatives both require wall-to-wall data (Sitch et al. 2008; Romijn et al. 2018; Herold et al. 2019). Thus, there is a strong need for going beyond local to regional AGB estimates to produce unified AGB density maps at the biome or global scale. Pantropical, temperate, boreal and global AGB maps have been proposed in the last decade, most often by extrapolating field-based and/or LiDAR-based AGB estimations through the use of global optical, RADAR and environmental datasets. In this section, we discuss the issues associated with large-scale AGB mapping and the promise of new methods and upcoming spaceborne missions.
5.1 The Problem of Interpolation
None of the RS systems available to date that display fair to good correlation with AGB (with no signal saturation) offers wall-to-wall coverage at very broad scales (i.e. regions, countries, continents). At best, these systems provide a fairly systematic sampling of the earth surface under the form of belt transects (e.g., ICESat/GLAS or the upcoming GEDI and MOLI missions). This condition implies that an interpolation step is necessary to produce a continuous AGB map from this discrete sampling. To that end, all published approaches have used or combined statistical interpolation (e.g., kriging or co-kriging with environmental drivers) and predictions from high- to medium-resolution spaceborne RS data to produce wall-to-wall coverage (e.g., Landsat, MODIS, QuickScat, ALOS PALSAR). For instance, Saatchi et al. (2011b) and Baccini et al. (2012) mapped AGB over the tropics by spatially interpolating discrete AGB estimates using Quick Scatterometer and/or MODIS data. Avitabile et al. (2016) then combined these two maps and used additional reference data (field plots and locally calibrated maps) in an attempt to generate an improved AGB map. Finally, the map from Santoro et al. (2018) was derived at 100-m resolution using a combination of maps based on ALOS PALSAR, ASAR data, Landsat data and ICESat GLAS transects, without any calibration from in situ data.
Thus, it is not surprising that strong inconsistencies have been reported among broad-scale AGB maps (Mitchard et al. 2013; Huang et al. 2015; Rodriguez-Veiga et al. 2017). For example, Mitchard et al. (2013) compared two widely used pantropical maps (Saatchi et al. 2011b, Baccini et al. 2012) and found fairly consistent estimates at the continental scale but large regional differences (> 100%) within continents. In Amazonia, a comparison with field estimates suggested that both pantropical maps failed to capture the AGB gradient across regions with some regions over- or underestimated by > 25% (Mitchard et al. 2014). However, without a systematic field sampling of AGB at a large scale, as done in national forest inventories (Kleinn 2017), it is difficult to validate whether a map captures the regional AGB gradients.
In Fig. 7, we report the result of a comparison of five existing pantropical/global AGB maps with six LiDAR-based AGB maps from Africa and French Guiana (Labriere et al. 2018). The results confirm that in dense forests (> 100 Mg ha−1), broad-scale maps display relatively weak (R2 ≤ 0.3) if any correlation at the pixel level with independent AGB data of improved quality. However, some maps exhibit much higher correlations than others considering the very coarse-resolution gradients (here the mean AGB per LiDAR site), with the R2 value for the map of Saatchi et al. (2011b) exceeding 0.8. Although this result is based on six observations and should be confirmed with more sites, it could suggest that broad-scale AGB variations may be successfully captured with such maps. These comparisons shed some light on the uncertainty associated with these AGB map estimates, but they also convey their own errors for validation or accuracy assessments for at least two reasons. First, the maps are all produced for different time periods, e.g., the Saatchi et al. (2011b) map is for ca. 2000, the Baccini et al. (2012) map is for ca. 2008 and Avitabile et al. (2016) combined these two maps using ground- and LiDAR-estimated AGB at different spatial scales and time periods. Second, the definition of forest and AGB may be different for each map. In Saatchi et al. (2011b), the biomass estimate at 1 km referred to the biomass of remaining forest in each 1 km grid cell, whereas in all other maps, the estimates referred to the mean biomass at the grid cell of the map (average of forest and non-forest biomass).
5.2 The Use of Environmental Proxies
One possible way to mitigate the aforementioned difficulties that hinder interpolation is to rely on spatialized environmental proxies to enhance the mapping. In the central Amazon, Saatchi et al. (2007) used a multivariate decision tree approach featuring both environmental variables (e.g., elevation, climate) and saturating RS metrics (e.g., Qscat_SV, NDVI) and found that AGB classes referring to values above 250 Mg ha−1 were mainly located in areas below 190 m elevation and of limited ruggedness (both assessed from SRTM). AGB is indeed known to covary with topography globally (Réjou-Méchain et al. 2014), but the strength and direction of the relationship strongly vary across sites (de Castilho et al. 2006; McEwan et al. 2011; Detto et al. 2013; Réjou-Méchain et al. 2015). Some studies successfully relied on spatialized climatic variables (Simard et al. 2011; Fayad et al. 2016) or geomorphological units (Guitet et al. 2015) to predict AGB or forest height in particular regional contexts. Thus, the ongoing development and improvement of global databases of environmental variables open avenues for geographically enlarging such approaches bearing in mind that relationships between AGB and the environment are intrinsically context-dependent. Indeed, AGB is largely driven by complex edaphic processes, such as soil structure (Gourlet-Fleury et al. 2011; Jucker et al. 2018b) or water table and bedrock depth (Emilio et al. 2013), which often vary at a relatively fine scale and for which continuous information is lacking. Furthermore, past and present anthropogenic activities, which are not randomly distributed according to abiotic conditions, may largely blur such environmental determinism at large scales. Therefore, calibration and prediction should be designed according to a relevant pre-stratification of the territory, accounting for anthropogenic activities, with sufficient reference data in each stratum.
5.3 Perspectives on Extrapolation
Upcoming RADAR satellite missions, such as NISAR in 2021 (Rosen et al. 2017) and BIOMASS in 2022 (Le Toan et al. 2011), should be pivotal in global-scale AGB mapping by providing wall-to-wall active measurements. While NISAR will be relevant for low AGB areas due to the use of an L-band saturating signal, BIOMASS will operate with a 50-m resolution P-band signal that is sensitive to large AGB values. (AGB products will, however, be delivered at 200-m resolution.) Preliminary analyses from P-band airborne missions revealed that polarimetric intensities (PolSAR) and polarimetric RADAR interferometry (PolInSAR) correlated well with AGB values up to 300 Mg ha−1 in temperate and tropical areas (ESA 2012). Above 300 Mg ha−1, such approaches failed to detect AGB variations, but tomographic approaches may significantly improve the retrieval of AGB up to 500 Mg ha−1 with no obvious saturation and a prediction error of only ca. 10% at a 1.5-ha resolution (Minh et al. 2014). A cross-validation performed between two sites of tropical forests in French Guiana suggested that tomographic models are transferable with an error within 20% at 1-ha resolution (Minh et al. 2016). However, these approaches rely on multiple acquisitions with different geometries over the same site, constituting a challenge for a spaceborne sensor and limiting the areas that may be covered in such a way. As a consequence, tomography will be implemented in only the first year of the mission; the three subsequent years will instead correspond to an interferometry phase. Beyond AGB mapping, the upcoming missions, including GEDI and MOLI, also constitute a unique opportunity to better understand the relationship between forest AGB, productivity and spatialized environmental proxies in tropical dense forests.
Finally, if future missions will considerably improve our ability to monitor AGB in space and time, the full potential of current RS products has not yet been completely assessed. For instance, using multi-angle MODIS observations to characterize the anisotropy of forests, de Moura et al. (2016) were able to generate metrics that displayed a good correlation (r2 > 0.5) with ALS-derived metrics in high AGB forests in the Brazilian Amazon, with no apparent saturation. This result adequately illustrates that there is also room for progress through the development of new methodologies to be used with the current RS products.
6 Radiative Transfer Modelling for Understanding Error Sources
A prerequisite to controlling the sources of errors in satellite-based assessments of forest AGB is to understand how the electromagnetic signal of interest interacts with forest components. Contrary to sea surface temperature, for instance, there is not yet any sensor that directly measures woody AGB per se. We thus use the variations in proximal measurements or indices to infer the AGB variations; however, these measures may be prone to interferences and artefacts that sometimes blur the true response of the forest AGB. Thus, radiative transfer (RT) modelling has become a standard approach in RS to analyse sensor sensitivity, and inverse modelling is used to measure the extent at which specific forest information is recoverable from the recorded signals. Some recent prominent examples can be found in studies that investigated the seasonal variation in leaf area in tropical forests (Morton et al. 2014; Wu et al. 2018) or texture retrieval of forest AGB from VHSR optical images (Ploton et al. 2017). Even crude modelling approaches may prove useful to untangle complex interactions in the sensor–atmosphere–scene–light source system (Barbier et al. 2011; Barbier and Couteron 2015), but a key point in such simulation studies is the way in which the complex 3D organization of tropical forest components can be represented. We will focus here on examples involving passive and active optical signals, but similar progress can be observed in RADAR studies (Villard 2009), which were historically pioneering (Ishimaru 1978).
6.1 An Increase in Model Realism
In the early days, geometric-optic models were used to understand the variation in directional reflectance within large pixels comprising a number of objects (e.g., trees) (Egbert 1977; Li and Strahler 1992; Roujean et al. 1992; Ni et al. 1999). Scene reflectance was computed as the sum of nominal reflectance of the components (e.g., shaded and lit surfaces) of a projection of simple, generally opaque, geometrical shapes within a scene. These efforts allowed for the derivation of analytical functions (kernels) that proved useful to fit actual directional reflectance variations with a small number of parameters (e.g., Lyapustin et al. 2011).
As computing power increased, it became possible to increase scene complexity, as well as to improve the level of detail in the descriptions of light–matter interactions. Discrete ray-tracing models based on Monte Carlo simulations of photon behaviour and (multiple) scattering within the scene have now become standard. This phenomenon is, for instance, evidenced by the periodic RAMI (RAdiation transfer Model Intercomparison) experiment, which aims at benchmarking and testing a number of existing models (Widlowski et al. 2013). Initially, scenes represented homogeneous parallel layers (Verhoef 1985; Myneni et al. 1989) of a turbid medium, in which the macroscopic transfer equation was statistically deduced from the microscopic distribution of elements, such as leaves or aerosols, within the cell. These elements were characterized by a density and angle distribution, as well as optical properties, which imply a probability of intercepting light rays depending on the incoming and outgoing directions. In addition to turbid elements, solid surfaces can be added in the form of triangle meshes to represent scene elements other than leaves (woody parts, roads, ground, building or water bodies). Triangles can be translucid and display Lambertian, specular or other optical properties. Progressively, spatial heterogeneity was introduced in the turbid medium distribution (Myneni 1991; Gastellu-Etchegorry et al. 1996; North 1996). Indeed, as opposed to most annual crops, crown shading is fundamental to understanding scene reflectance in forests. This fundamental need is even more pronounced when the spatial resolution of a sensor increases to a point where the pixels are smaller than the tree crowns, as biophysical descriptors can then be found in the texture features, and describing the spatial arrangement of reflectance values is crucial (Couteron et al. 2005; Barbier et al. 2011; Barbier and Couteron 2015). The degree of realism of a 3D description has evolved from sets of floating spheres or ellipsoids to more complex envelopes describing tree crown contours (Cescatti 1997; Widlowski et al. 2013). The dimensions and spatial distributions of these envelopes can generally be inferred from standard forestry measurements or allometric equations (Barbier et al. 2011).
6.2 Limits and Perspectives on RT Modelling
The current difficulties that need to be overcome to effectively derive forest mock-ups from TLS data are the same as those mentioned above for deriving AGB estimates from TLS. For instance, individual tree segmentation is required to allocate species-specific radiometric properties to individual trees. In addition, accurate segmentation of leaves from wood and proper estimation of leaf angle distribution are required to accurately simulate light–vegetation interactions (Antin et al. 2015). These steps become more difficult in the upper canopy as the density of the TLS point cloud decreases dramatically due to occlusion and increasing distance from the laser source. This difficulty may, however, be minimized through the fusion of TLS and UAV data. Finally, a better understanding of leaf demography and its integration into advanced modelling of leaf optical properties, combined with the improved description of non-photosynthetic tissues in LiDAR-derived 3D mockups, will increase the realism of simulations derived from physical modelling.
An alternative to producing a forest scene with a turbid description of leaves is to use an architectural model to simulate the tree branching structure and individual leaf distribution (De Reffye et al. 1995; Dauzat et al. 2001; Côté et al. 2011; Widlowski et al. 2013; Calders et al. 2018). Independently of the biological realism of such detailed scene descriptions, these models provide a means to assess the degradation of the transfer modelling outcome as the scene is progressively simplified in terms of representation (Jonckheere et al. 2006; Widlowski et al. 2013; Antin et al. 2015).
7 Conclusions and Perspectives
Because almost all steps of the processing chain rely on statistical models, the quality of calibration/validation data is crucial, yet often neglected. In this review, we first discussed common-sense procedures to drastically reduce the error levels at the field plot scale, such as improving georeferencing, increasing field plot size (1 ha or more) or developing allometric equations that better account for local ecological variations in plant composition and/or architecture. Terrestrial LiDAR scanning appears to be a promising avenue for individual-level and plot-level measurements of aboveground volumes. Combined with representative WD estimates that account for the intra- and inter-individual patterns of WD variations, robust predictions of AGB may be achievable. Furthermore, the combination of TLS and airborne LiDAR measurements in the upscaling scheme may significantly reduce the source of mismatches described in Sect. 3, allowing for the easy-to-perform control of coregistration and ensuring a continuum in the measured objects (Kükenbrink et al. 2017).
The extrapolation of local estimates to global scales should ideally be performed in successive steps, relying on intermediate, high-resolution data, such as VHSR spaceborne optical imagery, airborne or UAV-based LiDAR and (UAV-based) stereophotogrammetry. These data can be used to estimate AGB with an accuracy close to the accuracy of ground estimates, but they dramatically increase the representativeness of calibration data when earth observation data of decreased sensitivity (but also decreased cost) are to be used for regional to global predictions. The recent and rapid development of UAV systems, which currently fill the gap between terrestrial and airborne laser scanning systems, is particularly promising because these systems have the potential to bridge different spatial scales, from within-individual-tree to Earth observation pixels of medium resolution (e.g., MODIS).
In parallel, efforts should be expanded to better understand the interactions between different types of electromagnetic signals in varying contexts, e.g., RADAR, LiDAR and passive optical systems, to fully assess the potential of multisensor data fusion. These interactions may be largely blurred by numerous sources of noise and bias—instrumental, atmospheric, geometric, etc. Radiative Transfer (RT) modelling and the prospect to attain a high degree of realism in 3D stand modelling have made this approach a central tool for performing sensitivity analysis and designing signal correction algorithms. Assuming that 3D simulated stands will soon closely reflect reality, they may even be used to directly calibrate inversion models, bypassing many issues such as plot geolocation problems.
The last promise comes from the launch of new space missions (BIOMASS, GEDI, Icesat-2, MOLI, NISAR), which will improve our extrapolation capacities by providing both global coverage and signals of increased relevance. However, whether used individually or in fusion, the calibration of these products should rely on a pre-stratification of forested lands, with enough reference data in each stratum, as it is unlikely that a single universal model can be transferable across forest types and regions without biases.
To conclude, even high-quality Earth observation data currently do not meet the requirements of international environmental policies (e.g., IPCC GPG, REDD + , SDGs, see Herold et al. 2019). Understanding the sources of error and pathways to improve biomass estimation at all levels (as explained in this paper) is fundamental to implementing the best possible practices at each step, resulting in high-quality, consistent biomass estimates with uncertainties quantified to serve the various applications and users.
We gratefully thank the organizers of the Workshop held at ISSI Bern in November 2017 that was at the origin of this Special Issue. This review has been conducted under the project 3DForMod funded by ERA-GAS (ANR-17-EGAS-0002-01, NWO-3DForMod-5160957540) and has also benefited from the “Investissement d’Avenir” programs managed by Agence Nationale de la Recherche (CEBA, ref. ANR-10-LABX-25-01).
- Antin C, Grau E, Vincent G et al (2015) From leave scale to tree scale: which structural parameters influence a simulated full-waveform large-footprint LiDAR signal? SilviLaser 2015:110–112Google Scholar
- Arciniegas A, Prieto F, Brancheriau L, Lasaygues P (2014) Literature review of acoustic and ultrasonic tomography in standing trees. Trees 28:1559–1567Google Scholar
- Asner GP, Mascaro J (2014) Mapping tropical forest carbon: calibrating plot estimates to a simple LiDAR metric. Remote Sens Environ 140:614–624Google Scholar
- Baccini A, Asner GP (2013) Improving pantropical forest carbon maps with airborne LiDAR sampling. Carbon Manag 4:591–600Google Scholar
- Baker TR, Phillips OL, Malhi Y et al (2004) Variation in wood density determines spatial patterns in Amazonian forest biomass. Glob Change Biol 10:545–562Google Scholar
- Barbier N, Couteron P (2015) Attenuating the bidirectional texture variation of satellite images of tropical forest canopies. Remote Sens Environ 171:245–260Google Scholar
- Barbier N, Proisy C, Véga C et al (2011) Bidirectional texture function of high resolution optical images of tropical forest: an approach using LiDAR hillshade simulations. Remote Sens Environ 115:167–179Google Scholar
- Bastin J-F, Fayolle A, Tarelkin Y et al (2015a) Wood specific gravity variations and biomass of central african tree species: the simple choice of the outer wood. PLoS ONE 10:e0142146Google Scholar
- Bastin J-F, Barbier N, Réjou-Méchain M et al (2015b) Seeing Central African forests through their largest trees. Sci Rep 5:13156Google Scholar
- Bauwens S, Bartholomeus H, Calders K, Lejeune P (2016) Forest inventory with terrestrial LiDAR: a comparison of static and hand-held mobile laser scanning. Forests 7:127Google Scholar
- Béland M, Baldocchi DD, Widlowski J-L et al (2014) On seeing the wood from the leaves and the role of voxel size in determining leaf area distribution of forests with terrestrial LiDAR. Agric For Meteorol 184:82–97Google Scholar
- Blanchard E, Birnbaum P, Proisy C et al (2015) Prédire la structure des forêts tropicales humides calédoniennes: analyse texturale de la canopée sur des images Pléiades. Rev Fr Photogrammétrie Télédétection 209:141–147Google Scholar
- Bouvet A, Mermoz S, Le Toan T et al (2018) An above-ground biomass map of African savannahs and woodlands at 25 m resolution derived from ALOS PALSAR. Remote Sens Environ 206:156–173Google Scholar
- Bouvier M, Durrieu S, Fournier RA, Renaud J-P (2015) Generalizing predictive models of forest inventory attributes using an area-based approach with airborne LiDAR data. Remote Sens Environ 156:322–334Google Scholar
- Brede B, Lau A, Bartholomeus HM, Kooistra L (2017) Comparing RIEGL RiCOPTER UAV LiDAR derived canopy height and DBH with terrestrial LiDAR. Sensors 17:2371Google Scholar
- Calders K, Origo N, Burt A et al (2018) Realistic forest stand reconstruction from terrestrial LiDAR for radiative transfer modelling. Remote Sens 10:933Google Scholar
- Cescatti A (1997) Modelling the radiative transfer in discontinuous canopies of asymmetric crowns. I. Model structure and algorithms. Ecol Model 101:263–274Google Scholar
- Chave J, Condit R, Aguilar S et al (2004) Error propagation and scaling for tropical forest biomass estimates. Philos Trans R Soc Lond Ser B-Biol Sci 359:409–420Google Scholar
- Chave J, Coomes D, Jansen S et al (2009) Towards a worldwide wood economics spectrum. Ecol Lett 12:351–366Google Scholar
- Chen Q, Laurin GV, Valentini R (2015) Uncertainty of remotely sensed aboveground biomass over an African tropical forest: propagating errors from trees to plots to pixels. Remote Sens Environ 160:134–143Google Scholar
- Clark DA (2002) Are tropical forests an important carbon sink? Reanalysis of the long-term plot data. Ecol Appl 12:3–7. https://doi.org/10.1890/1051-0761(2002)012%5b0003:atfaic%5d2.0.co;2 Google Scholar
- Condit R (1998) Tropical forest census plots: methods and results from Barro Colorado Island, Panama and a comparison with other plots. Springer, BerlinGoogle Scholar
- Condit R, Ashton PS, Baker P et al (2000) Spatial patterns in the distribution of tropical tree species. Science 288(5470):1414–1418Google Scholar
- Condit R, Lao S, Singh A et al (2014) Data and database standards for permanent forest plots in a global network. For Ecol Manag 316:21–31Google Scholar
- Couteron P, Pelissier R, Nicolini EA, Paget D (2005) Predicting tropical forest stand structure parameters from Fourier transform of very high-resolution remotely sensed canopy images. J Appl Ecol 42:1121–1128Google Scholar
- Dauzat J, Rapidel B, Berger A (2001) Simulation of leaf transpiration and sap flow in virtual plants: model description and application to a coffee plantation in Costa Rica. Agric For Meteorol 109:143–160Google Scholar
- de Moura YM, Hilker T, Goncalves FG et al (2016) Scaling estimates of vegetation structure in Amazonian tropical forests using multi-angle MODIS observations. Int J Appl Earth Obs Geoinf 52:580–590Google Scholar
- De Reffye P, Houllier F, Blaise F et al (1995) A model simulating above-and below-ground tree architecture with agroforestry applications. Agrofor Syst 30:175–197Google Scholar
- de Souza Pereira FR, Kampel M, Gomes Soares ML et al (2018) Reducing uncertainty in mapping of mangrove aboveground biomass using airborne discrete return LiDAR data. Remote Sens 10:637Google Scholar
- Egbert DD (1977) A practical method for correcting bidirectional reflectance variations. In: LARS symposia, p 203Google Scholar
- ESA (2012) Report for mission selection: biomass, ESA SP-1324/1 (3 volume series). European Space Agency Noordwijk, The NetherlandsGoogle Scholar
- Fayad I, Baghdadi N, Guitet S et al (2016) Aboveground biomass mapping in French Guiana by combining remote sensing, forest inventories and environmental data. Int J Appl Earth Obs Geoinf 52:502–514Google Scholar
- Feldpausch TR, Banin L, Phillips OL et al (2011) Height–diameter allometry of tropical forest trees. Biogeosciences 8:1081–1106Google Scholar
- Féret J-B, Asner GP (2014) Mapping tropical forest canopy diversity using high-fidelity imaging spectroscopy. Ecol Appl 24:1289–1296Google Scholar
- Ferraz A, Saatchi S, Mallet C, Meyer V (2016) LiDAR detection of individual tree size in tropical forests. Remote Sens Environ 183:318–333Google Scholar
- Féret JB, Gitelson AA, Noble SD, Jacquemoud S (2017) PROSPECT-D: towards modeling leaf optical properties through a complete lifecycle. Remote Sens Environ 193:204–215Google Scholar
- Frazer GW, Wulder MA, Niemann KO (2005) Simulation and quantification of the fine-scale spatial pattern and heterogeneity of forest canopy structure: a lacunarity-based method designed for analysis of continuous canopy heights. For Ecol Manag 214:65–90Google Scholar
- Frazer GW, Magnussen S, Wulder MA, Niemann KO (2011) Simulated impact of sample plot size and co-registration error on the accuracy and uncertainty of LiDAR-derived estimates of forest stand biomass. Remote Sens Environ 115:636–649Google Scholar
- Fuller WA (1987) Measurement error models. Wiley, New YorkGoogle Scholar
- Gao S, Wang X, Wiemann MC et al (2017) A critical analysis of methods for rapid and nondestructive determination of wood density in standing trees. Ann For Sci 74:27Google Scholar
- Gastellu-Etchegorry J-P, Demarez V, Pinel V, Zagolski F (1996) Modeling radiative transfer in heterogeneous 3-D vegetation canopies. Remote Sens Environ 58:131–156Google Scholar
- Gastellu-Etchegorry J-P, Yin T, Lauret N et al (2015) Discrete anisotropic radiative transfer (DART 5) for modeling airborne and satellite spectroradiometer and LiDAR acquisitions of natural and urban landscapes. Remote Sens 7:1667–1701Google Scholar
- Gobakken T, Naesset E (2009) Assessing effects of positioning errors and sample plot size on biophysical stand properties derived from airborne laser scanner data. Can J For Res 39:1036–1052Google Scholar
- Gonzalez de Tanago J, Lau A, Bartholomeus H et al (2018) Estimation of above-ground biomass of large tropical trees with terrestrial LiDAR. Methods Ecol Evol 9:223–234Google Scholar
- Grau E, Durrieu S, Fournier R et al (2017) Estimation of 3D vegetation density with terrestrial laser scanning data using voxels. A sensitivity analysis of influencing parameters. Remote Sens Environ 191:373–388Google Scholar
- Gregoire TG, Næsset E, McRoberts RE et al (2016) Statistical rigor in LiDAR-assisted estimation of aboveground forest biomass. Remote Sens Environ 173:98–108Google Scholar
- Guitet S, Pélissier R, Brunaux O et al (2015) Geomorphological landscape features explain floristic patterns in French Guiana rainforest. Biodivers Conserv 24:1215–1237Google Scholar
- Guitet S, Sabatier D, Brunaux O et al (2018) Disturbance regimes drive the diversity of regional floristic pools across Guianan rainforest landscapes. Sci Rep 8:3872Google Scholar
- Hajj ME, Baghdadi N, Fayad I et al (2017) Interest of integrating spaceborne LiDAR data to improve the estimation of biomass in high biomass forested areas. Remote Sens 9:213Google Scholar
- Hansen MC, Potapov PV, Moore R et al (2013) High-resolution global maps of 21st-century forest cover change. Science 342:850–853Google Scholar
- Henry M, Besnard A, Asante WA et al (2010) Wood density, phytomass variations within and among trees, and allometric equations in a tropical rainforest of Africa. For Ecol Manag 260:1375–1388Google Scholar
- Huang W, Swatantran A, Johnson K et al (2015) Local discrepancies in continental scale biomass maps: a case study over forested and non-forested landscapes in Maryland, USA. Carbon Balance Manag 10:19Google Scholar
- Hunter MO, Keller M, Victoria D, Morton DC (2013) Tree height and tropical forest biomass estimation. Biogeosciences 10:8385–8399Google Scholar
- Inglada J, Vadon H (2005) Fine registration of SPOT5 and Envisat/ASAR images and ortho-image production: a fully automatic approach. In: Proceedings 2005 IEEE international geoscience and remote sensing symposium, 2005. IGARSS'05. IEEE, Vol. 5, pp 3510–3512Google Scholar
- Ishimaru A (1978) Wave propagation and scattering in random media, vol 2. Academic press, New York, pp 336–393Google Scholar
- Johnson CE, Barton CC (2004) Where in the world are my field plots? Using GPS effectively in environmental field studies. Front Ecol Environ 2:475–482. https://doi.org/10.1890/1540-9295(2004)002%5b0475:witwam%5d2.0.co;2 Google Scholar
- Jonckheere I, Nackaerts K, Muys B et al (2006) A fractal dimension-based modelling approach for studying the effect of leaf distribution on LAI retrieval in forest canopies. Ecol Model 197:179–195Google Scholar
- Jucker T, Asner GP, Dalponte M et al (2017a) A regional model for estimating the aboveground carbon density of Borneo’s tropical forests from airborne laser scanning. arXiv Prepr arXiv170509242Google Scholar
- Jucker T, Caspersen J, Chave J et al (2017b) Allometric equations for integrating remote sensing imagery into forest monitoring programmes. Glob Change Biol 23:177–190Google Scholar
- Jucker T, Asner GP, Dalponte M et al (2018a) Estimating aboveground carbon density and its uncertainty in Borneo’s structurally complex tropical forests using airborne laser scanning. Biogeosciences 15:3811–3830Google Scholar
- Jucker T, Bongalov B, Burslem DF et al (2018b) Topography shapes the structure, composition and function of tropical forest landscapes. Ecol Lett 21:989–1000Google Scholar
- Justice CO, Giglio L, Korontzi S et al (2002) The MODIS fire products. Remote Sens Environ 83:244–262Google Scholar
- Kearsley E, De Haulleville T, Hufkens K et al (2013) Conventional tree height–diameter relationships significantly overestimate aboveground carbon stocks in the Central Congo Basin. Nat Commun 4:2269Google Scholar
- Kellner JR, Asner GP (2009) Convergent structural responses of tropical forests to diverse disturbance regimes. Ecol Lett 12:887–897Google Scholar
- Kennedy RE, Yang Z, Cohen WB (2010) Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr—temporal segmentation algorithms. Remote Sens Environ 114:2897–2910Google Scholar
- Kleinn C (2017) The renaissance of National Forest Inventories (NFIs) in the context of the international conventions—a discussion paper on context, background and justification of NFIs. Pesqui Florest Bras 37:369–379Google Scholar
- Labriere N, Tao S, Chave J et al (2018) In situ reference datasets from the TropiSAR and AfriSAR campaigns in support of upcoming spaceborne biomass missions. IEEE J Sel Top Appl Earth Obs Remote Sens 99:1–11Google Scholar
- Lau A, Bentley LP, Martius C et al (2018) Quantifying branch architecture of tropical trees using terrestrial LiDAR and 3D modelling. Trees 32(5):1219–1231Google Scholar
- Lefsky MA, Cohen WB, Parker GG, Harding DJ (2002) LiDAR remote sensing for ecosystem studies LiDAR, an emerging remote sensing technology that directly measures the three-dimensional distribution of plant canopies, can accurately estimate vegetation structural attributes and should be of particular interest to forest, landscape, and global ecologists. Bioscience 52:19–30Google Scholar
- Leitold V, Morton DC, Longo M et al (2018) El Niño drought increased canopy turnover in Amazon forests. New Phytol 219:959–971Google Scholar
- Li X, Strahler AH (1992) Geometric-optical bidirectional reflectance modeling of the discrete crown vegetation canopy: effect of crown shape and mutual shadowing. IEEE Trans Geosci Remote Sens 30:276–292Google Scholar
- Lindenmayer DB, Cunningham RB, Tanton MT et al (1991) Characteristics of hollow-bearing trees occupied by arboreal marsupials in the montane ash forests of the Central Highlands of Victoria, south-east Australia. For Ecol Manag 40:289–308Google Scholar
- Liu J-Y, Zheng Z, Xu X et al (2018) Abundance and distribution of cavity trees and the effect of topography on cavity presence in a tropical rainforest, southwestern China. Can J For Res 48:1058–1066Google Scholar
- Longo M, Keller M, dos-Santos MN et al (2016) Aboveground biomass variability across intact and degraded forests in the Brazilian Amazon. Glob Biogeochem Cycles 30:1639–1660Google Scholar
- Lopez-Gonzalez G, Lewis SL, Burkitt M, Phillips OL (2011) ForestPlots.net: a web application and research tool to manage and analyse tropical forest plot data. J Veg Sci 22:610–613Google Scholar
- Lyapustin A, Martonchik J, Wang Y et al (2011) Multiangle implementation of atmospheric correction (MAIAC): 1. Radiative transfer basis and look-up tables. J Geophys Res Atmos 116:1–9Google Scholar
- Ma L, Zheng G, Eitel JU et al (2016) Improved salient feature-based approach for automatically separating photosynthetic and nonphotosynthetic components within terrestrial LiDAR point cloud data of forest canopies. IEEE Trans Geosci Remote Sens 54:679–696Google Scholar
- Malenovskỳ Z, Martin E, Homolová L et al (2008) Influence of woody elements of a Norway spruce canopy on nadir reflectance simulated by the DART model at very high spatial resolution. Remote Sens Environ 112:1–18Google Scholar
- McRoberts RE, Westfall JA (2013) Effects of uncertainty in model predictions of individual tree volume on large area volume estimates. For Sci 60(1):34–42Google Scholar
- Minh DHT, Le Toan T, Rocca F et al (2014) Relating P-band synthetic aperture radar tomography to tropical forest biomass. IEEE Trans Geosci Remote Sens 52:967–979Google Scholar
- Minh DHT, Le Toan T, Rocca F et al (2016) SAR tomography for the retrieval of forest biomass and height: cross-validation at two tropical forest sites in French Guiana. Remote Sens Environ 175:138–147Google Scholar
- Mitchard ET, Saatchi SS, Baccini A et al (2013) Uncertainty in the spatial distribution of tropical forest biomass: a comparison of pan-tropical maps. Carbon Balance Manage 8:10Google Scholar
- Momo Takoudjou S, Ploton P, Sonké B et al (2018) Using terrestrial laser scanning data to estimate large tropical trees biomass and calibrate allometric models: a comparison with traditional destructive approach. Methods Ecol Evol 9:905–916Google Scholar
- Morsdorf F, Eck C, Zgraggen C et al (2017) UAV-based LiDAR acquisition for the derivation of high-resolution forest and ground information. Lead Edge 36:566–570Google Scholar
- Moundounga Mavouroulou Q, Ngomanda A, Engone Obiang NL et al (2014) How to improve allometric equations to estimate forest biomass stocks? Some hints from a central African forest. Can J For Res 44:685–691Google Scholar
- Myneni RB (1991) Modeling radiative transfer and photosynthesis in three-dimensional vegetation canopies. Agric For Meteorol 55:323–344Google Scholar
- Myneni RB, Ross J, Asrar G (1989) A review on the theory of photon transport in leaf canopies. Agric For Meteorol 45:1–153Google Scholar
- Ni W, Li X, Woodcock CE et al (1999) An analytical hybrid GORT model for bidirectional reflectance over discontinuous plant canopies. IEEE Trans Geosci Remote Sens 37:987–999Google Scholar
- North PR (1996) Three-dimensional forest light interaction model using a Monte Carlo method. IEEE Trans Geosci Remote Sens 34:946–956Google Scholar
- Pargal S, Fararoda R, Rajashekar G et al (2017) Inverting aboveground biomass–canopy texture relationships in a landscape of Forest mosaic in the Western Ghats of India using very high resolution Cartosat imagery. Remote Sens 9:228Google Scholar
- Pearson TR, Brown S, Murray L, Sidman G (2017) Greenhouse gas emissions from tropical forest degradation: an underestimated source. Carbon Balance Manag 12:3Google Scholar
- Phillips OL, Baker TR, Brienen R, Feldpausch TR (2009) Field manual for plot establishment and remeasurement. https://www.forestplots.net/upload/ManualsEnglish/RAINFOR_field_manual_EN.pdf
- Ploton P, Pélissier R, Proisy C et al (2012) Assessing aboveground tropical forest biomass using Google Earth canopy images. Ecol Appl 22:993–1003Google Scholar
- Ploton P, Pélissier R, Barbier N et al (2013) Canopy texture analysis for large-scale assessments of tropical forest stand structure and biomass. In: Devy S, Ganesh T, Lowman MD (eds) Treetops at risk. Springer, Berlin, pp 237–245Google Scholar
- Ploton P, Barbier N, Momo ST et al (2016) Closing a gap in tropical forest biomass estimation: taking crown mass variation into account in pantropical allometries. Biogeosciences 13:1571–1585Google Scholar
- Ploton P, Barbier N, Couteron P et al (2017) Toward a general tropical forest biomass prediction model from very high resolution optical satellite images. Remote Sens Environ 200:140–153Google Scholar
- Puliti S, Ørka HO, Gobakken T, Næsset E (2015) Inventory of small forest areas using an unmanned aerial system. Remote Sens 7:9632–9654Google Scholar
- Raumonen P, Kaasalainen M, Åkerblom M et al (2013) Fast automatic precision tree models from terrestrial laser scanner data. Remote Sens 5:491–520Google Scholar
- Réjou-Méchain M, Muller-Landau HC, Detto M et al (2014) Local spatial structure of forest biomass and its consequences for remote sensing of carbon stocks. Biogeosciences 11:6827–6840Google Scholar
- Réjou-Méchain M, Tymen B, Blanc L et al (2015) Using repeated small-footprint LiDAR maps to infer spatial variation and dynamics of a high-biomass neotropical forest. Remote Sens Environ 169:93–101Google Scholar
- Réjou-Méchain M, Tanguy A, Piponiot C et al (2017) BIOMASS: an r package for estimating above-ground biomass and its uncertainty in tropical forests. Methods Ecol Evol 8:1163–1167Google Scholar
- Robinson C, Saatchi S, Neumann M, Gillespie T (2013) Impacts of spatial variability on aboveground biomass estimation from L-band radar in a temperate forest. Remote Sens 5:1001–1023Google Scholar
- Rocchini D, Luque S, Pettorelli N et al (2018) Measuring β-diversity by remote sensing: a challenge for biodiversity monitoring. Methods Ecol Evol 9:1787–1798Google Scholar
- Rodrigues WA, Valle RC (1964) Ocorrência de troncos ocos em mata de baixio da regiao de Manaus, 16th edn. Publicacao. Botanica - Instituto Nacional de Pesquisa da Amazonia (Brazil), ManausGoogle Scholar
- Rodriguez-Veiga P, Wheeler J, Louis V et al (2017) Quantifying forest biomass carbon stocks from space. Curr For Rep 3:1–18Google Scholar
- Romijn E, De Sy V, Herold M et al (2018) Independent data for transparent monitoring of greenhouse gas emissions from the land use sector—what do stakeholders think and need? Environ Sci Policy 85:101–112Google Scholar
- Roşca S, Suomalainen J, Bartholomeus H, Herold M (2018) Comparing terrestrial laser scanning and unmanned aerial vehicle structure from motion to assess top of canopy structure in tropical forests. Interface Focus 8:20170038Google Scholar
- Rosen P, Hensley S, Shaffer S et al (2017) The NASA-ISRO SAR (NISAR) mission dual-band radar instrument preliminary design. In: 2017 IEEE international geoscience and remote sensing symposium (IGARSS), pp 3832–3835Google Scholar
- Roujean J-L, Leroy M, Deschamps P-Y (1992) A bidirectional reflectance model of the Earth’s surface for the correction of remote sensing data. J Geophys Res Atmos 97:20455–20468Google Scholar
- Saatchi S, Mascaro J, Xu L et al (2015) Seeing the forest beyond the trees. Glob Ecol Biogeogr 24:606–610Google Scholar
- Santoro M, Cartus O, Mermoz S et al (2018) A detailed portrait of the forest aboveground biomass pool for the year 2010 obtained from multiple remote sensing observations. In: EGU general assembly conference abstracts, p 18932Google Scholar
- Schlund M, von Poncet F, Kuntz S et al (2015) TanDEM-X data for aboveground biomass retrieval in a tropical peat swamp forest. Remote Sens Environ 158:255–266Google Scholar
- Schneider FD, Yin T, Gastellu-Etchegorry J et al (2014) At-sensor radiance simulation for airborne imaging spectroscopy. In: 2014 6th workshop on hyperspectral image and signal processing: evolution in remote sensing (WHISPERS), pp 1–4Google Scholar
- Simard M, Pinto N, Fisher JB, Baccini A (2011) Mapping forest canopy height globally with spaceborne LiDAR. J Geophys Res Biogeosci 116:1–12Google Scholar
- Singh M, Malhi Y, Bhagwat S (2014) Biomass estimation of mixed forest landscape using a Fourier transform texture-based approach on very-high-resolution optical satellite imagery. Int J Remote Sens 35:3331–3349Google Scholar
- Sitch S, Huntingford C, Gedney N et al (2008) Evaluation of the terrestrial carbon cycle, future plant geography and climate-carbon cycle feedbacks using five dynamic global vegetation models (DGVMs). Glob Change Biol 14:2015–2039Google Scholar
- Solberg S, May J, Bogren W et al (2018) Interferometric SAR DEMs for forest change in Uganda 2000–2012. Remote Sens 10:228Google Scholar
- Steininger MK (2000) Satellite estimation of tropical secondary forest above-ground biomass: data from Brazil and Bolivia. Int J Remote Sens 21:1139–1157Google Scholar
- St-Onge B, Vega C, Fournier RA, Hu Y (2008) Mapping canopy height using a combination of digital stereo-photogrammetry and LiDAR. Int J Remote Sens 29:3343–3364Google Scholar
- Sullivan MJ, Lewis SL, Hubau W et al (2018) Field methods for sampling tree height for tropical forest biomass estimation. Methods Ecol Evol 9:1179–1189Google Scholar
- Sun G, Ranson KJ (2000) Modeling LiDAR returns from forest canopies. IEEE Trans Geosci Remote Sens 38:2617–2626Google Scholar
- Swenson NG, Enquist BJ (2008) The relationship between stem and branch wood specific gravity and the ability of each measure to predict leaf area. Am J Bot 95:516–519Google Scholar
- Trochta J, Krŭček M, Vrška T, Král K (2017) 3D Forest: an application for descriptions of three-dimensional forest structures using terrestrial LiDAR. PLoS ONE 12:e0176871Google Scholar
- Tymen B, Vincent G, Courtois EA et al (2017) Quantifying micro-environmental variation in tropical rainforest understory at landscape scale by combining airborne LiDAR scanning and a sensor network. Ann For Sci 74:32Google Scholar
- Verhoef W (1985) Earth observation modeling based on layer scattering matrices. Remote Sens Environ 17:165–178Google Scholar
- Vieilledent G, Vaudry R, Andriamanohisoa SF et al (2012) A universal approach to estimate biomass and carbon stock in tropical forests using generic allometric models. Ecol Appl 22:572–583Google Scholar
- Villard L (2009) Forward and inverse modeling of synthetic aperture radar in the bistatic configuration: applications in forest remote sensing. Ph.D. thesis, ONERAISAE-Universite Paul SabatierGoogle Scholar
- Villard L, Le Toan T (2015) Relating P-band SAR intensity to biomass for tropical dense forests in hilly terrain: γ 0 or t 0? IEEE J Sel Top Appl Earth Obs Remote Sens 8:214–223Google Scholar
- Vincent G, Caron F, Sabatier D, Blanc L (2012a) LiDAR shows that higher forests have more slender trees. Bois For Trop 314:51–56Google Scholar
- Vincent G, Sabatier D, Rutishauser E (2014) Revisiting a universal airborne light detection and ranging approach for tropical forest carbon mapping: scaling-up from tree to stand to landscape. Oecologia 175:439–443Google Scholar
- Vincent G, Antin C, Laurans M et al (2017) Mapping plant area index of tropical evergreen forest by airborne laser scanning. A cross-validation study using LAI2200 optical sensor. Remote Sens Environ 198:254–266Google Scholar
- Wassenberg M, Chiu H-S, Guo W, Spiecker H (2015) Analysis of wood density profiles of tree stems: incorporating vertical variations to optimize wood sampling strategies for density and biomass estimations. Trees 29:551–561Google Scholar
- Williamson GB, Wiemann MC (2010) Measuring wood specific gravity… correctly. Am J Bot 97:519–524Google Scholar
- Wu X, Liu H, Li X et al (2018) Differentiating drought legacy effects on vegetation growth over the temperate Northern Hemisphere. Glob Change Biol 24:504–516Google Scholar
- Xu L, Saatchi SS, Yang Y et al (2016) Performance of non-parametric algorithms for spatial mapping of tropical forest structure. Carbon Balance Manag 11:18Google Scholar
- Xu L, Saatchi SS, Shapiro A et al (2017) Spatial distribution of carbon stored in forests of the Democratic Republic of Congo. Sci Rep 7:15030Google Scholar