Estimating Forest Above-Ground Biomass in Central Amazonia Using Polarimetric Attributes of ALOS/PALSAR Images

Abstract

Polarimetric synthetic aperture radar (SAR) images are essential to understand forest structure and plan forest inventories with the purpose of natural resource management and environmental conservation efforts. We developed a method for estimating above-ground biomass (AGB) from power and phase-radar attributes in L-band images. The model was based on the variables “P_v” (from Freeman–Durden decomposition) and “σ°_HH”, complemented by the attributes of Touzi decomposition “α_S2”, “τ_m”, “ Φ_S3”, and “ Φ_S2”. The analyses demonstrated the contribution of volumetric, multiple, and direct scattering resulting from the interaction between the signal and the random structure of canopies and their forest biomass. The proposed model had good predictive capacity and a positive correlation (R² = 0.67 and = 0.81, respectively), with Syx = 56.9 Mg ha⁻¹ and a low average estimation error of 7.5% at R² = 0.81 in the validation. An additional exploratory analysis of the parallel polarimetric responses did not reveal a defined pattern for the different phytophysiognomies—although all indicated a predominance of multiple and/or volumetric scattering. This fact can be related to the floristic and structural variation in the primary forest units, the degree of human intervention in legal logging, and the differences among succession stages.

1. Introduction

According to the Food and Agriculture Organization [1], the world’s land cover by forests has changed from 31.6 to 30.6% during the 1990–2020 period, corresponding to an absolute area of 40 million km².

The Amazon Forest is the largest and most continuous area of tropical forest in the world in terms of coverage and biodiversity. This biome deserves special attention due to the increasing impact caused by deforestation, which has increased since 2012, achieving a rate of 10,851 km² yr⁻¹ [2]—a much higher value than that established by the international agreement (ENREDD+—National REDD+ Strategy and Signatory Countries to UNFCCC—United Nations Framework Convention on Climate Change) for 2020 (3925 km² yr⁻¹) [3].

Forests have an important role in the terrestrial carbon cycle [4]. For this reason, the monitoring of AGB (from local to global scales) has become a challenge for studies on climate change [5], as it requires the use of data from orbital sensor systems to increase the efficiency in the acquisition and periodicity of information, maximizing coverage and minimizing time and cost of forest biomass estimation [5,6].

In tropical regions, the use of radar systems allows for acquiring data independent of the time of the day, cloud coverage, and atmospheric conditions [7]. SAR in L- and P-band datasets are the most appropriate sensor for forest studies, particularly the estimation of biomass and forest biophysical variables [8], allowing us to understand the interaction between emitted energy and tree structure at different ecological stages [9]. These types of sensors not only provide valuable information on the spatial distribution and quantification of forest biomass and carbon [10] but also have promising results [11] because of the frequency and polarization that they operate [12] as well as the improved spatial and temporal resolution of the system [13].

Some studies have considered the variation in biomass due to different forest typologies [14], floristic and structural diversity [15,16], and local geomorphometric conditions [17,18]. Polarimetric analysis techniques have been used for biomass estimation [19,20], including interferometric approaches [21,22,23,24] employing interferometric height measurements and polarimetric SAR interferometry (PoLinSAR) [25,26], which have resulted in lower errors for biomass prediction [27] under different phytophysiognomic conditions [14], in addition to a reduced backscattering signal saturation [28], especially for tropical dense forests [29].

Although many studies have been devoted to improving global biomass estimation [30,31,32,33], there is much uncertainty in these forecasts on account of the low spatial resolution of the instruments used (wide mode) and the absence of sample units or their discontinuity over the continents. In some cases, there may be an undersized dimension of these sample units in relation to the pixel size used, in addition to rainfall variation on a global scale, impairing the quality of the models generated under these conditions [31].

Estimations made on a local scale tend to be more accurate [34,35,36], mainly with regard to the monitoring of local constraints, such as the dimension and spatial location of sample units, and the rainfall pattern (which can affect the image quality) due to dielectric constant impacts [36] in the images. Moreover, the use of fine-mode images and correction of terrain variations [37] lead to better estimates of forest inventory data, which are required for the management and conservation of forest biomass at local scale. In this context, biomass maps at finer scales play an important role in decision-making, which requires more accurate spatial information, and they can be used to assess and adjust broader-scale biomass maps.

Remote sensing is used to measure, for instance, the height of trees or the number of tree stems per unit area to biomass estimation which is inferred by field measurements establishing a connection between tree plant biomass and satellite-based remote sensing measurements [38], especially the most accurate of all mode acquisitions radar images using the phase, of better spatial resolution in the along the line narrower of sight, covering local surveys. Thus, combining local data from independent sources, such as remote sensing images allows a robust assessment of total plant biomass [39], on account of this, the present research also contemplates local biomass predictions, which are more accurate than global estimates in tropical zones [40], which present high uncertainty and can overestimate lower biomass sites and underestimate tropical areas with high biomass [30,32].

The relevance of local research, which includes predictions of biomass and radar image with phase preservation to provide reliable results, is evidenced. For instance, for presenting the Touzi decomposition in the final biomass model, the parameter considered invariant to polarization base, ideal for natural targets described by [41,42] as an essential parameter for improving the results in biomass models in the Amazon Várzea Wetlands [40] and Amazon terra firme exclusively in primary forest sites [43]. However, the potential of this type of decomposition was not tested yet for biomass estimation of tropical forest sites in different successional stages.

For these reasons, our research introduces beyond a wide range of coherent and incoherent attributes in the constructed model, the parallel polarimetric responses as a differential to other research in similar regions as a way to improve the complete understanding of the relation signal-target [44] in a wide range of forest growth stages and managements for understanding the relationship between forest structural variations and respective biomass patterns in our model and consequently being as a reliable and comparable basis to the improvement of global biomass evaluations in tropical zones [30].

In this context, we developed a methodology to estimate forest biomass by using L-band polarimetric data for the Amazon Forest typologies with different floristic and structural diversity in primary and secondary forests in different successional stages as well as forests undergoing logging operations.

2. Materials and Methods

2.1. Study Area

The study area has approximately 109,930 ha encompassing the Tapajós National Forest (TNF) (a protection area) and its surroundings (3°23′4.89″ S–2°56′45.34″ S; 54°46′51.57″ W–55°5′47.38″ W), in the north of Pará state, Brazil (Figure 1). The climate in the region is tropical wet—Am in the Köppen classification [45]. The average annual temperature is 24.8 °C, and the average annual rainfall is 2100 mm, with a dry season from July to September (with rainfall below 60 mm) (data acquired from a meteorological station in Belterra, Pará, Brazil, located about 90 km northeast of the study area (02°38′ S; 54°05′ W)) [46].

The local topography varies from flat to strongly undulated and is divided into two very different morphostructural units: the Tapajós-Xingu plateau (altimetric height of 120−170 m) and the lower plateau of the middle Amazonas River (altimetric height of ~100 m). The soil is predominantly dystrophic yellow latosol, associated with red-yellow argisols [47].

An analysis of the digital elevation model of the shuttle radar topography mission (SRTM) for the study area showed that the altitudes vary between 80 and 200 m, covering both morphostructural units.

The study area is covered by remnants of dense and open ombrophilous forests as well as forest areas of sustainable logging. Neighboring the TNF and along the federal highway BR-163 are extensive areas of mechanized agriculture and pasture, besides a historical settlement of small-scale landowners where family farming predominates, including areas of secondary succession at different stages of regeneration. More details about the forest classes and their physical characteristics can be found in the studies by [19,48].

2.2. Acquisition of SAR and Inventory Data

The SAR image was obtained on 21 October 2006, from the Advanced Land Observing Satellite-1/Phased Array L-band SAR-1 (ALOS/PALSAR) sensor. The data were acquired in full polarimetric mode, ascending orbit, an incidence angle of 24.31°, and a resolution of 3.56 × 9.37 m (line and pixel spacing, respectively).

According to the data from the meteorological station near the study area, two days before the image acquisition date, the precipitation was 8.5 mm. This acquisition date corresponded to the end of the dry period and the beginning of the rainy season (i.e., October), when no increases were recorded in the dielectric constant values and variations in the radiometric contrast for the studied forest typologies.

To validate this study, the floristic-structural characterization of the landscape was described, and the AGB was estimated from forest inventories of 41 field plots (4 units with dimensions 10 × 25 m (Maltese cross format)—primary forest and selective logging; 100 × 100 m—selective logging; 10 × 100 m—intermediate and advanced secondary succession; 10 × 50 m—initial secondary succession) units (33 training samples and 8 validation samples) from August to September 2006.

The coordinates of the vertices of the sample polygons were obtained in the UTM zone (21S projection, WGS84 datum) using the navigation Global Positioning System (GPS) with horizontal position error between 3 and 6 m. Subsequently, the relative positioning errors were evaluated, and the variations were compensated by a buffer as a function of the horizontal deviations related to this process in the x and y directions of the image. Each sample unit was inventoried based on its geographic orientation to the north using a compass and the distance of the current tree in the x- and y-axes (in meters) from the reference point using a measuring tape (lower left corner).

The canopy openness fraction was obtained along the center line of all sample units using hemispherical photographs. This type of measurement provides information about the remaining trees and their growth at each measurement point as well as their structural differences [49], being considered fundamental for the proposed analysis. Through the z-adjusted Wald–Wolfowitz test [50], it was possible to evaluate the differences among the phytophysiognomies.

To characterize the forest landscape, we used sample units with varying dimensions, depending on the floristic structural variations of the primary and secondary typologies, as adopted by [51].

In each sample unit, the total height (TH, measured in meters) and commercial height (CH) of each tree were employed to correlate the polarization scattering mechanisms and the polarimetric attributes that constituted the final biomass model. The measurement of the diameter at breast height (DBH, measured in centimeters), determined for different growth stages and/or forest classes, corresponded to the limit of inclusion of trees in the sample.

During the inventory process, the identification of species was done by a botanist. Details regarding the identification, checking, confirmation, and grouping of species and their respective families are presented in [52].

The aboveground biomass (AGB) of each tree was estimated using allometric equations. For primary forests, the allometric equation proposed by [53] (Equation (1)) was used to estimate AGB in Mg (megagrams or metric tons). The same equation was applied to forest samples with logging and secondary succession at an advanced stage (age > 25 years). For the initial secondary and intermediate stages of succession, the allometric equation developed by [54] (Equation (2)) was used. Both equations were later weighted by area (to Mg ha⁻¹) according to the sample size used. The equations are below, where in both equations, DBH represents the diameter at breast height and TH represents the total height of the tree. In Equation (2), the natural logarithm (ln) is used to calculate the AGB estimate, and the exponent function (exp) is used to obtain the final AGB estimate:

$$AGB\left(Mg\right)=0.044\times {\left({\left(DBH\right)}^{²}\times TH\right)}^{0.9719}$$

(1)

$$In AGB\left(Mg\right)=-2.17+ 1.02 ln{\left(DBH\right)}^2 + 0.39 ln\left(TH\right)$$

(2)

2.3. SAR Polarimetric Data Processing and Analysis

Through single-look complex (SLC) image processing, it was possible to convert the scattering matrix [S] into a covariance matrix [C], while coherence [T] was estimated based on multilook processing with seven azimuth looks for each 1-range look (Figure 2). This process produced images with approximate resolutions of 22.85 m and 24.92 m (azimuth and range, respectively). Afterward, a modified Lee filter [55] with a 5 × 5 pixel window was applied to reduce speckle noise.

The SAR parameters for each forest sample unit were estimated from the image [T] in slant range projection. This procedure was a demand to preserve and extract phase information, and consequently obtain coherent attributes from the element of high spatial resolution in the slant range. To accomplish this, we conducted a process to convert the geographic coordinates (WGS-84 cartographic reference system, UTM 21S datum) from the vertices of the sample units into coordinates of the multilook image—the so-called geocoding, which was performed on the Envi’s polarimetric module [56].

The SAR coherent parameters analyzed were polarimetric coherence and polarimetric phase difference [57]; entropy, anisotropy, and mean alpha angle [58]; magnitude, angle orientation phase, and ellipticity [53,54]; and volumetric scattering, double bounce, and surface component [59]. Further, the SAR incoherent parameters selected were backscatter coefficient, parallel and cross-polarization ratio, and total power [57]; biomass index, canopy structure, and volumetric scattering [60]; and intensity values. The processing was carried out using PolSARpro software version 4.0, developed by ESA (European Space Agency).

To convert the signal to backscatter coefficient sigma naught (σ°), Equation (3) [61] was applied using PCI Geomatica software version 10.1, as follows:

σ° [dB] = 10 log₁₀ <I² + Q²> + cf − A,

(3)

where I and Q are the real and imaginary parts of the SLC product (level 1.1), i.e., <I² + Q²> = Amplitude, cf is the radiometric calibration factor (−83), and A is the conversion factor (32).

Image orthorectification was performed to correct positioning errors resulting from altimetry variation using the first-degree polynomial model and resample them by the nearest neighbor method [62]. This procedure corrected the distortions caused by the relief due to the geometry of the radar sight and allowed us to use the results of biomass model generation and compare them with the values extracted from the forest inventory.

The relationship between backscatter and forest structure was investigated through multivariate analysis, with AGB being selected as the dependent variable.

2.4. Modeling of Above-Ground Tree Biomass

To evaluate the potential of ALOS/PALSAR data to estimate AGB, a multiple regression analysis was applied. A total of 33 forest plots were studied, including six plots of primary forest, five of secondary succession, eight of intermediate secondary succession, seven of initial secondary succession, and seven of forest with selective logging. For validation, eight plots of different forest phytophysiognomies were used, extrapolating more than the minimum required amount (20%) of total samples, in agreement with the statistical recommendations [50].

The multivariate analysis was preceded by an exploratory data analysis to evaluate the correlations between explanatory variables and the independent variable [50].

The choice of the final model was determined by the best subsets of explanatory variables (best subset method) using selection techniques based on the coefficient of determination (R²), adjusted coefficient of determination (R² adjusted), and Mallow’s Cp criterion, which were analyzed in conjunction with the mean square of the residues (MSR) [50]. After this step, the previously selected model was refined by evaluating the presence of interaction effects (bivariate terms), multicollinearity diagnosis (variance inflation factors –VIF), outlier analysis (Cook’s distance), and residue analysis. After generating the best-fit model (estimated biomass) and comparing the obtained data with the measured biomass (observed in the field), the validation step was conducted by calculating the uncertainty according to the standard deviation error (Syx = √MSR) and its respective mean estimation error, R², and Pearson’s correlation coefficient (r) [50].

3. Results

3.1. SAR, Inventory Data, and Scattering Mechanisms

The values of the structural parameters demonstrated a decreasing trend of biomass, which in turn increased the density of individuals and the fraction of canopy openness because of the persistent increase of human interference (Appendix A). According to the basic concept, a parallel polarimetric response with similar σ_HH and σ_VV values (Figure 3a) was observed in PF2; in the unit (ISS31; Figure 3b), higher values of σ were obtained in the linear polarization in HH (χ = 0° and ψ = ±0°).

The initial secondary succession showed higher canopy openness than the primary forest, as demonstrated by the Wald–Wolfowitz runs test (z-adjusted = 1.811422; p-value = 0.070076), indicating a greater contribution of soil backscattering and a lower contribution of forest structure (mean DBH = 8.87 cm). Further, the primary forest exhibited the opposite behavior due to its high density and smaller canopy openness in this successional stage, suggesting a greater contribution of thinner trunks and thicker branches, which was reflected in the initial rotation of this parallel polarimetric response. For the PF2 and ISS31 units, the density of individuals (404 and 1780 ind ha⁻¹, respectively) was higher than that in the other PF and ISS units (385 and 1193 ind ha⁻¹ on average, respectively), which may justify their high pedestal values (

Table 1. Behavior of parallel polarimetric responses and predominant scattering mechanisms. This parameter was analyzed graphically as a function of ellipticity (χ) and orientation (ψ) angles. Thus, the sum of the parallel polarization polarimetric syntheses responses of many individual scatterers for a given pixel can identify the dominant scattering mechanisms and the degree of randomness of the backscatter signal using the pedestal height as a parameter [59,60]. It is worth mentioning that different types of scattering show different values of pedestal height [63]. Forest typologies were associated with the predominant scattering mechanism, and consequently, the parallel polarimetric response, and their pedestal height, as proposed by [44].

Typology	Scattering Mechanism	HH Response to VV	Average Pedestal Height (σ)
Primary forest (PF)
PF1, PF20, PF21, PF29	multiple and volumetric scattering	HH > VV	0.142; 0.121; 0.137 and 0.125
PF2, PF9	multiple scattering	HH < VV or HH ≈ VV	0.251; 0.151
Mean			0.155
Selective logging (SL)
SL25	multiple scattering	HH < VV or HH ≈ VV	0.16
SL15, SL36, SL37	multiple and volumetric scattering	HH > VV	0.199; 0.155; 0.122,
SL38, SL40, SL41			0.147; 0.141; 0.148
Mean			0.153
Advanced secondary succession (ASS)
ASS18, ASS27	multiple and volumetric scattering	HH > VV	0.142; 0.097
ASS33, ASS34, ASS35	multiple scattering	HH < VV or HH ≈ VV	0.116; 0.147; 0.144
Mean			0.129
Intermediate secondary succession (IntSS)
IntSS3, IntSS8,	multiple scattering	HH < VV or HH ≈ VV	0.141; 0.157;
IntSS10, IntSS17			0.126; 0.088
IntSS7, IntSS12,	multiple and volumetric scattering	HH > VV	0.107; 0.094,
IntSS22, IntSS26			0.141; 0.099
Mean			0.119
Initial secondary succession (ISS)
ISS4, ISS5, ISS6	multiple scattering	HH < VV or HH ≈ VV	0.161; 0.089; 0.109
ISS28			0.133
ISS24, ISS31, ISS32	multiple and volumetric scattering	HH > VV	0.112; 0.121; 0.117
Mean			0.127

). Therefore, the structural variations are reflected in changes in the parallel polarimetric response behavior [44].

The intermediate secondary succession unit exhibited a behavior considered atypical (IntSS8; Figure 3c) since the σ values occurred in a circular polarization to the right (χ = −45°), with a vertical orientation angle (ψ = ±90°). This can be attributed to its diametrical distribution and differentiated vertical structure in relation to the other units in this succession stage. The lowest number of individuals (22) with higher average height (TH = 21.26 m) and mean diameter ($\mathrm{D}\mathrm{B}\mathrm{H}$ = 23.18 cm), of which eight emerged from the pioneer species of irregular Jacaranda copaiba D. Don, contrasts with the great majority of individuals (133) with lower mean diameter ($\mathrm{D}\mathrm{B}\mathrm{H}$ = 8.6 cm) and average height ($\mathrm{T}\mathrm{H}$ = 11.42 m) belonging to the lower strata (Figure 3d), as evidenced by the distinct behavior of parallel polarimetric responses, the floristic-structural variations caused by anthropic activities and the site variations, which led to a greater pedestal height in this successional stage.

3.2. Above-Ground Tree Biomass

The model did not indicate multicollinearity problems (

Table 2. Statistical parameters derived from AGB based on incoherent and coherent polarimetric SAR attributes (σ°_HH is the sigma naught in HH polarization; P_v is the volume scattering component of the Freeman–Durden decomposition; α_S2 is the Touzi magnitude of the 2nd component; Φ_S2 is the Touzi phase of the 2nd component; Φ_S3 is the Touzi phase of the 3rd component; and τ_m is the Touzi ellipticity of the main component).

Variable	βk	Sk	t *	p *	VIF
Constant	−1221.37	442.3	−2.76	0.010
σ°HH	−70.31	26.818	−2.62	0.014	4.43
Pv	1064.65	327.2	3.25	0.003	4.91
αS2	6.28	2.916	2.15	0.041	2.07
ΦS2	−2.42	1.264	−1.91	0.067	1.16
ΦS3	3.44	1.294	2.66	0.013	1.57
τm	6.05	3.474	1.74	0.094	1.29

(*) Generated with a 0.05% confidence level. Note: R² = 0.67; R² (adj) = 0.33; r = 0.81; PRESS = 190,351.8; SSR = 84,385.63; p-value = 0.009424 (highly significant value).

), which means that there was a non-significant correlation between the explanatory variables in the regression model (lower values of variance inflation value—VIF). The relationship between the measured AGB and the parameters derived from the SAR data was established using their respective regression coefficients and significance from the t-test. Such a relationship was found to be significant for all selected variables.

The variables HH polarization (σ°_HH) and P_v were the ones that most contributed to the prediction of AGB (higher values of βk), whereas the Touzi components had similar but lower values. A multiple regression model capable of explaining the heterogeneity of the scattering elements present in forest phytophysiognomies was generated (Equation (4); Figure 4). This model was based on the statistical criteria used (R² = 0.67; r = 0.81; and p = 0.0094).

AGB = −1221.37 − 70.31 (σ°_HH) + 1064.65 (P_v) + 6.28 (α_S2) − 2.42 (Φ_S2) + 3.44 (Φ_S3) + 6.05 (τ_m),

(4)

where AGB is the above-ground biomass; σ°_HH is the backscatter coefficient in the HH polarization (in dB); P_v is the Freeman–Durden volumetric scattering component (in dB); α_S2 is the Touzi magnitude of the intermediate scattering component (2nd component) (in degree); Φ_S2 is the phase of the intermediate scattering component (2nd component) (in degree); Φ_S3 is the phase of the minor scattering component (3rd component) (in degree); and τ_m is the ellipticity of the dominant scattering component (in degree) generated by the Touzi decomposition.

In this model, backscatter in σ°_HH had a strong negative correlation with AGB, indicating an inverse relationship between these scattering elements in comparison with vertically arranged tree trunks, which most contributed to backscatter due to their total amount of AGB.

The Touzi phase of the intermediate spreading component (Φ_S2), which constituted part of the model, had a moderately negative correlation with AGB, with average values of 4.05°, being classified as direct scattering.

When generating the model, the strong negative correlation with σ°_HH and the moderately negative correlation with Φ_S2 indicated a strong correlation of AGB with tree trunks of preferably vertical layout and volumetric scattering, respectively. In this case, the greatest contribution of biomass in primary forest areas and old logging occurs in larger diameter classes and consequently, higher height classes and components of the upper and emerging canopy (classes of trees with TH > 30 m, as shown in Figure 5). Although the values of biomass were mostly calculated for the lower strata, the highest results were obtained for the upper strata, being significantly increased to 170 and 210 Mg ha⁻¹ for selective logging and primary forests, respectively. These values confirm the use of the L-band and the attributes of σ°_HH and mainly the direct spreading from the upper and emerging stratum of more developed individuals resulting from the contribution of Φ_S2.

The P_v variable had the highest positive correlation with AGB, highlighting the importance of this type of scattering for the model formulation, confirmed by the results of parallel polarimetric responses, mainly in primary and selective logging generated according to the methodology established in [44]; for tropical forests results in thick branches of high structural complexity and the random structure of the different canopy heights found, these were simulated according to [59], a cloud of randomly arranged dipoles. This pattern was also confirmed by [64] in similar areas of primary and secondary forests, P_v component also contributes more than 90% when compared to the surface and double bounce backscattering components generated.

The Touzi magnitude of the intermediate component (2nd component) also showed a positive correlation with AGB (α_S2 = 60.47°) with a predominance of multiple scattering. The ellipticity of the dominant scattering component generated by the Touzi decomposition contributed to the intermediate magnitude for the AGB model in a similar way. The interaction between the L-band and the complex structure of the study area was confirmed by the asymmetry of natural targets (τ_m = −22.50°), with the spreading of the propeller type being oriented to the right. Finally, although small compared to Φ_S2, the positive correlation of the Touzi phase of the minor scattering component (3rd component—Φ_S3~50°) demonstrated the importance of spreading from multiple interactions.

During validation, the estimated regression model values were similar to the inventory data (Figure 6a)—even though in some cases they generated errors over 15%, especially for the PF19 and ASS30 units, (Figure 6b). The lack of adjustment for the referred primary forest unit had a deviation of 17.7% from its observed value and a residue of 40.93 Mg ha⁻¹ (Figure 6b), while for the ASS30 unit, these values were 25.17% from and 33.8 Mg ha⁻¹, respectively.

In the model validation process, the uncertainty of the estimation of AGB showed reasonable efficacy, with a standard deviation error Syx = 56.9 Mg ha⁻¹. For this reason, the model was considered with good predictive capacity since the average error of the low estimate was only 7.45%. In addition, the predictive capacity of this model showed satisfactory adjustment between field biomass data and those generated by the remote sensing model (R² = 0.81 and r = 0.90).

4. Discussion

When parallel polarimetric responses are similar in HH and VV, as in the case of PF2, they are close to the theoretical response of a trihedral corner reflector [65], with a predominant double-bounce scattering [64]. Similar behavior was observed for selective logging units in an area close to our study area [66] and in various forest species in a temperate forest region [67]. These patterns were detected on surfaces considered rough and under different angles of incidence [68], mainly at high SAR frequencies.

In primary forest and selective logging units, it was possible to observe the maximum response of σ in HH in linear polarizations (χ = 0°), with a preferably horizontal orientation (ψ = 0°). These parallel polarimetric responses were related to the occurrence of more compact upper canopies of lower canopy openness (Appendix A), behavior resulting from horizontally oriented thick branches [19]. This suggests multiple scattering and consequently volumetric variations among the pixels, as reported by [44] during an extensive analysis conducted in tropical forest areas using the L-band.

For the initial secondary succession unit (ISS34), the parallel polarimetric response behavior was due to the small contribution of tree trunks, the less developed diameter in the individuals, and the high canopy openness in relation to primary forest areas. With greater soil exposure, this phytophysiognomy tends to have a more homogeneous canopy, with most of the thick branches of the remaining trees comprising the emergent stratum. They are horizontally oriented and preponderant for this configuration, with higher values of σ in linear polarizations (χ = 0°) and HH orientation (ψ = 0°) [64]. Similar backscatter behavior was predominantly found in heavily and frequently burned forests using ALOS/PALSAR [69]. However, the authors suggested a greater contribution of horizontally arranged constituents, such as fallen trunks or branches in areas severely affected by fire.

For PF2, ISS31, and IntSS8, the pedestal height was higher than for the other units of the same forest physiognomy and succession since this variable is directly related to the high density of individuals found [70], as evidenced by the forest inventory (Appendix A). In the forest, the height of the relatively high pedestal can be related to the presence of multiple and dissimilar scatterers, typical of targets dominated by volumetric scattering originating from multiple interactions [62,66]. This corroborates the results found by [71] for softwood forest classes mainly composed of white pine and black spruce and hardwood forest with red oak species. According to these authors, these forest species reached the highest accuracy in their classification due to their high pedestal.

The forest structural variation caused by either site conditions or human interference was responsible for the atypical behavior found in the intermediate secondary succession unit (IntSS8). This demonstrates a parallel polarimetric response that is similar to the theoretical response of a right-oriented helix [65], indicating depolarization of the incident wave due to forest density at the intermediate successional stage, resulting in multiple scattering on account of randomly scattering elements. These results are in agreement with the behavior observed in a tropical forest region in Guyana [44] and the northern Amazon [69]. When analyzing only primary forests, [72] found similar behavior in units with undulated and high undulated relief arising from topographic effects.

The lack of a tropical forest standard for parallel polarimetric responses can be explained by the high physiognomic and structural complexity of these sites [64], resulting from the combined influence of the density and spatial distribution of trees, trunks, twigs, and branches. This was also verified by [19] when analyzing different stages of forest succession and physiognomy. However, the authors reported a predominance of multiple and volumetric scattering.

The spatial arrangement of AGB density in tropical forests is very complex, thus having a strong influence on the radar response because of its scattering mechanisms [20]. Among other factors, the increased number of strata is responsible for increasing both the CO₂ concentration in the forest canopy and the structural integrity of the forest, with at least three strata being identified in intact tropical [73]. This makes it difficult to generate more accurate predictive models, justifying the use of polarimetry [43] proposed by [40], who found the best results in L-band full-polarimetric data in comparison with single/dual-pol SAR data in floodplain forests of the Amazon basin.

Although tropical and subtropical forests are amongst the most structurally complex ecosystems in the world (with a three-dimensional arrangement of forest canopy elements, such as leaves, branches, and trunks), the greatest contribution comes from tree trunks (75%), which are responsible for a greater accumulation of AGB [74]. In the present study, the highest values of biomass were observed for trees that compose the upper strata (trees with TH > 30 m), especially in primary forests and forests under selective logging due to their larger basal area. Astiani et al. [75] verified this behavior in tropical lowland forests with lower intensity of degradation.

The volumetric component generated in the Freeman–Durden (P_v) decomposition, which in this analysis had the highest positive correlation with AGB, is associated with the good fit to the model for the volumetric component in the upland tropical forest for the L-band made by the developers of the proposed method [59], evidencing the importance of volumetric scattering.

Several studies on forest structure and generation of volumetric and biomass models have recognized the significance of this attribute [65,66,74]. According to [9], the volumetric component shows a greater variability of radiometric responses for different typologies in the same area.

The values of α_S2, which had a positive correlation with AGB, confirm the structural complexity of the different classes, demonstrating the importance of multiple scattering [41,42]. This is corroborated by the ellipticity values of the dominant scattering component (τ_m), with a similar contribution to α_S2.

In this study, a positive correlation was found with AGB (although lower than the other correlations of the same trend), where the variable Φ_S3, a component of the biomass model, proved the contribution of multiple interactions [42]. This arrangement, which has a high number of randomly distributed elements, also controls volume backscattering in the resolution cell [20]. According to [65], in forest environments the forest structure is considered a cloud of randomly oriented dipoles, simulating the condition of the different strata. Other studies also proved that the scattering is influenced by dipoles [59,67], confirming that the penetration of the C-band into the tree canopy is low and that the waves do not often reach the trunks.

In particular, τ_m is fundamental for the complementarity of the Freeman volume scattering contribution to the total power (P_v), besides providing information analogous to the type of scattering generated by angle α [58] that aim to remove the ambiguity of this type of scattering, particularly for asymmetric scattering [42]. Additionally, since this parameter is invariant to rotation in natural targets, it is more suitable for the modeling of biomass in tropical forests, as verified by [43,66] in a region similar to the study area. In this regard, [16] analyzed the historical use of a secondary tropical forest area in the stage of regeneration after abandonment and observed a high correlation with angular decompositions, with the Touzi target phase angle being the highest correlation (Φ_S) found near the TNF in Pará state.

Despite having the lowest adherence to AGB, the second component of the Touzi phase (Φ_S2) showed values related to direct backscattering, thus providing reliable information about the target mechanisms involved [41,42,76]. In their study, [35] also demonstrated the relative importance of Φ_S2 in biomass models generated in China’s temperate forest areas dominated by Pinus yunnanensis, Larix gmelinii, and Betula platyphylla species. This scattering mechanism found diverges from that reported by [40], who described it as a predictor that provides information about phase difference between trihedral and dihedral scattering, but is congruent with [70], who demonstrated the importance of the 2nd component of the Touzi phase, relating it to double-bounce scattering and considering it an important feature for estimating the AGB in the Amazon Forest.

Several studies have been conducted in various forest types using different SAR systems and techniques, resulting in different levels of precision in the generation of predictive models (

Table 3. Review of studies using different SAR techniques and their respective forest types in biomass estimation.

SAR Sensor	Biomass Saturation (Mg ha−1)	Precision (R2)	RMSE (Mg ha−1)	Technique	Forest Type	Author (s)
ALOS-1/ PALSAR-1	~210	0.67	56.9	Full polarimetric (coherent and incoherent attributes)	Tropical forest in different successional stages	This study
ALOS-1/ PALSAR-1	300	0.53–0.55	92–94	Michigan Microwave Canopy Scattering model (MIMICS-I)	Deciduous mixed forest	[77]
ALOS-1/ PALSAR-1	≥100	not provided	20	Polarised data (HV polarization)	Eleven forest types	[15]
	~80	not provided	20		Fresh water swamp forests
	~250	not provided	20		Needleleaf forests
AIRSAR (L band)	≤150	0.42–0.91	23.8	Full polarimetric	Tropical wet forest	[14]
(P-band)	˂300	0.68–0.91	22.6	Full polarimetric
(C-band)	≤150 + 40%	0.79	29	PolinSAR (L + InSAR height index)
	˂300 + 20%	0.93	15.1	PolinSAR (P + InSAR height index)
TerraSAR-X	~280	0.73	58	Radargrammetric height	Spruce forest	[23]
Tandem-X	~300	0.66	65	Interferometric height	Spruce forest	[22]
ALOS-2/PALSAR-2 and RADARSAT-2	>120	0.562 (site I) 0.461 (site II)	16.02 (site I) 28.46 (site II)	Polarimetric attributes	China’s temperate forest areas	[35]
ALOS-1/ PALSAR-1	~210	44.70%	54.3	Full polarimetric (coherent attributes)	Dense and open ombrophilous forest	[43]
	~210	44.90%	50.6	Full polarimetric (coherent attributes + cosine factor (terrain)
	~210	79.40%	33.2	Full polarimetric (coherent attributes + elevation + slope)
ALOS-2/ PALSAR-2	150	0.90	14.24	Full polarimetric (coherent and incoherent attributes) + terrain factors + RF + PCA + ridge regression (to solve the problem of collinearity)	Mixed forest of coniferous and broad-leaved trees	[78]
ALOS-1/PALSAR -1; Envisar ASAR; GLAS	147	57.1%	10.6	WDBEF (Wood Density Biomass Expansion Factor) + sigma nought ground + sigma nought vegetation layers = GSV (growing stock volume (m3 ha−1)	Global estimation	[31]
GlobBiomass 2010; CCI Biomass 2017 v.1 (SAR backscattered intensity images); Baccini 2000 (LiDAR footprints + Landsat Reflectance); GEOCARBON 2007–2010 map (combining a refined pantropical map and radar backscatter intensity)	200–250	not provided	76.75 (in average)	Global biomass maps and NFIs	Global estimation	[30]

).

A study conducted by [77] to retrieve biomass in deciduous mixed forest employed the Michigan microwave canopy scattering model (MIMICS-I) using ALOS/PALSAR. Although these authors found a higher saturation point for biomass, they showed a higher RMSE than that observed in our analysis, under more complex structural conditions.

Although considered a model with good predictive capacity and similar biomass saturation point, when compared to [43], who developed a study in a similar area, this model showed high precision and low error—even though it was done in areas with topographic variation, which as a rule changes the backscatter values to the same class type in the scene, depending on the slope. The insertion of the digital elevation model as a variable into the biomass model confirms its importance due to the causal relationship with ecophysiological aspects that condition the growth of forest communities and minimize the terrain effects on the radar signal, thus improving the predictive capacity of the model.

The inclusion of interferometric techniques in models that use multi-polarized data and previously established polarimetric attributes (with phase preservation) increases precision, reduces errors, and minimizes signal saturation for high biomass values. The results reported by [14,79], who conducted a study in a tropical rainforest region, prove that the insertion of interferometric height into the proposed model leads to higher results than the techniques hitherto established [22,23]. Forest biomass using X-band interferometric and radargrammetric techniques that successfully predicted models with high levels of biomass saturation point and good precision for forests with less floristic-structural complexity, such as Abeto forests, was studied by [22,23]. However, they presented superior errors in comparison with the developed models using polarimetric and PolInSAR techniques in tropical forests, which have a more complex structure.

Global forest models based on L-band SAR backscatter developed by [15] demonstrated that the relationship between L-band backscatter and AGB can be significantly different, depending on the forest type and environmental effect, consequently requiring multiple algorithms to map AGB from global time series of radar satellite observations.

Therefore, it can be concluded that the structural complexity, the relief variations, and the techniques employed directly influence the significance of the models generated. Further details can be found in the studies by [14,15,22,23,43].

The lack of adjustment for some sample units already mentioned is linked to the transition area between the plateau forest and the lowland forest (slope forest). To reduce this limitation, Liu et al. [78] developed a new approach using L-band and inserting coherent and incoherent attributes alongside terrain factors into their model. In addition, to improve biomass estimation these authors used random forest (RF), principal component analysis (PCA), and ridge regression methods to solve the multicollinearity problem between variables in the model. Although conducted in a forest with less floristic and structural complexity than the present research, this study showed better results. According to [80], elevation and topography variations affect soil composition and microclimate. This is directly related to the increased spatial arrangement of the forest structure caused by human intervention and, consequently, the density of AGB in tropical forests, in accordance with the data obtained herein.

The low values of Syx found in the biomass stock of the study are associated with the use of L-band radar phase information, which can be used by polarimetry, interferometry, or a combination of both since these are preferred techniques for the evaluation of biomass in heterogeneous environments of high biophysical complexity [81].

The efficacy of the model is proven through its comparison with studies conducted in less complex forest environments, such as forest plantations, which obtained standard error values similar to those found in the present study [5]. The authors concluded that their model had a weak correlation for regions of high biomass due to the high values of biomass and the multilayer structure, which were responsible for signal saturation. The variables selected in the present study, which led to a grouping of different stages of forest succession with primary forest and a sustainable management regime, differ from those in previous studies carried out in the TNF and surrounding areas.

These studies also include the works by [43], who used full polarimetric images exclusively of forests in different relief variations for extracting coherent attributes and geomorphological variables (L-band) for biomass estimates, and [66], who analyzed data from airborne polarimetric images (L-band) for estimating the volume in forest areas under logging regime.

The proven accuracy of our analysis in generating a biomass model in an environment with a complex forest structure is similar to the study by [35], who used the C- and L-bands and proved that the prediction models with Touzi (for the C- and L-bands) and Yamaguchi attributes (only for the C-band) were fundamental to determine the double-bounce and volumetric scattering mechanisms for both sites. However, even in a temperate forest, the accuracy of such work was lower than that found in the present study.

Although less precise than local models, information on global biomass models directly supports public policies aimed at reducing emissions from deforestation and degradation (REDD+) by quantifying national carbon stocks. In this context, this initiative has a market-based approach to reducing greenhouse gas emissions from deforestation and forest degradation [82]. Finally, by better understanding the carbon stock patterns and their dynamics it is possible to gain increased knowledge of forest biomass pools, helping constrain Earth system models [31].

In this perspective, [31] used Envisar ASAR (C-band) and ALOS/PALSAR (L-band) images, and GLAS (Lidar) data to create a global biomass map, capturing the spatial pattern and magnitude of the biomass, except for regional uncertainties in high carbon stocks with AGB > 250 Mg ha⁻¹, while underestimating the entire pantropical area. However, they found an error similar to that observed in local biomass estimation studies.

In the same sense, [30] used four global biomass maps generated with radar backscatter intensity and auxiliary data, including Lidar Footprints and Landsat Reflectance, and compared them to their framework in order to assess their accuracy and uncertainty. The results showed that biomass saturation reached medium ranges with high bias error, which was underestimated to 150 Mg ha⁻¹ in the pantropical forest region, mainly due to the low plot dimension and resolution of SAR images.

Despite their crucial relevance to the REDD+ policy through the monitoring of global biomass and carbon stocks, these approaches produce less precise biomass models since they underestimate the tropical rainforest zones in comparison with studies at local scales.

5. Conclusions

The polarimetric technique used provided good results for modeling heterogeneous phytophysiognomies with random scattering elements typical of the rainforest. This was evidenced by the low standard error of estimate, high correlation coefficient, and resulting coefficient of determination. The explanatory variable with the highest positive correlation in the composition of the model (P_v) proved the importance of the different forest canopy strata, which composed the phytophysiognomies analyzed and caused the volumetric scattering. The inverse relationship between the incoherent attribute σ°_HH and the biomass model indicated the secondary importance of the horizontally arranged scattering elements and the greater relevance of the vertically arranged tree trunks. The introduction of Touzi coherent attributes, α_S2, τ_m, and Φ_S3, was fundamental for the development of the AGB model, evidencing the secondary importance of multiple scattering, which caused the depolarization of the wave and a moderate inverse relationship between the direct scattering verified and Φ_S2 due to the increased penetration capability of the L-band. The lack of a typical pattern for parallel polarization responses in different classes and stages of forest succession in the rainforest can be attributed to the density of individuals and their complex structure, the number of forest canopies, and the intensity of anthropogenic intervention, leading to a variation of patterns and indicating multiple and volumetric scattering. We prove that local estimates are more accurate than global biomass estimates in the tropical zone, serving as a basis for more accurate global models in these regions. Based on the results, efforts should be directed towards the application of polarimetry, interferometry, as well as multi-frequency SAR approach, which can improve the accuracy of biomass estimates in tropical forests and contribute to product calibration and validation to reduce the uncertainty of global biomass maps.