Correcting Underestimation and Overestimation in PolInSAR
Forest Canopy Height Estimation Using Microwave
Penetration Depth
Hongbin Luo 1,2 , Cairong Yue 1,2,*, Ning Wang 1,2, Guangfei Luo 1,2 and Si Chen 1
1 College of Forestry, Southwest Forestry University, Kunming 650224, China
2 Forestry 3S Engineering Technology Research Center, Southwest Forestry University, Kunming 650224, China
* Correspondence:
cryue@swfu.edu.cn; Tel.: +86-13577178073
Abstract: PolInSAR is an active remote sensing technique that is widely used for forest canopy height
estimation, with the random volume over ground (RVoG) model being the most classic and effective
forest canopy height inversion approach. However, penetration of microwave energy into the forest
often leads to a downward shift of the canopy phase center, which leads to model underestimation
of the forest canopy height. In addition, in the case of sparse and low forests, the canopy height is
overestimated, owing to the large ground-to-volume amplitude ratio in the RVoG model and severe
temporal decorrelation effects. To solve this problem, in this study, we conducted an experiment on
forest canopy height estimation with the RVoG model using L-band multi-baseline fully polarized
PolInSAR data obtained from the Lope and Pongara test areas of the AfriSAR project. We also
propose various RVoG model error correction methods based on penetration depth by analyzing
the model’s causes of underestimation and overestimation. The results show that: (1
) In tall forest
areas, there is a general underestimation of canopy height, and the value of this underestimation
correlates strongly with the penetration depth, whereas in low forest areas, there is an overestimation
of canopy height owing to severe temporal decorrelation; in this instance, overestimation can also
be corrected by the penetration depth. (2) Based on the reference height RH100, we used training
sample iterations to determine the correction thresholds to correct low canopy overestimation and
tall canopy underestimation; by applying these thresholds, the inversion error of the RVoG model can
be improved to some extent. The corrected R2
increased from 0.775 to 0.856, and the RMSE decreased
from 7.748 m to 6.240 m in the Lope test area. (3) The results obtained using the infinite-depth volume
condition p-value as the correction threshold were significantly better than the correction results
for the reference height, with the corrected R2 value increasing from 0.775 to 0.914 and the RMSE
decreasing from 7.748 m to 4.796 m.
(4) Because p-values require a true height input, we extended the
application scale of the method by predicting p-values as correction thresholds via machine learning
methods and polarized interference features; accordingly, the corrected R2
increased from 0.775 to
0.845, and the RMSE decreased from 7.748 m to 6.422 m. The same pattern was obtained for the
Pongara test area. Overall, the findings of this study strongly suggest that it is effective and feasible
to use penetration depth to correct for RVoG model errors.
Keywords: forest canopy height; penetration depth; overestimation; underestimation; PolInSAR
1. Introduction
Forest canopy height is one of the most fundamental forest structure parameters and
represents an essential indicator to characterize both forest growth and carbon sink capacity [1,2]. Traditional manual methods to measure forest height are not only laborious and
time-consuming but are also limited to obtaining information from specific plots, making
it difficult to achieve large-scale and long-term observations; in addition, topographic
and climatic limitations can hinder large regional-scale surveys, resulting in gaps in forest canopy height monitoring coverage [3,4]. Currently, remote sensing techniques are
Remote Sens. 2022, 14, 6145.
https://doi.org/10.3390/rs14236145 https://www.mdpi.com/journal/remotesensing
Remote Sens. 2022, 14, 6145 2 of 27
primarily used to measure forest canopy height information at large regional scales, and
commonly applied remote sensing methods include optical, LiDAR, and synthetic aperture
radar remote sensing [5]. Optical remote sensing is less sensitive to the vertical structure of
forests, prone to saturation, and affected by weather. Benefits of LiDAR include that it is
applicable under all weather conditions and can actively describe the 3D vertical structure
information of vegetation. However, the observation scale and time scale of this approach
are limited by the high operational costs and time-consuming nature of aerial LiDAR surveys. In contrast, synthetic aperture radar (SAR) avoids the abovementioned shortcomings
in measuring forest height information over large areas via mechanistic models or empirical
or semi-empirical models, and its observation time scale is relatively large [6].
Polarimetric interferometric SAR (PolInSAR) is an active remote sensing technique
that is widely used for forest canopy height inversion [4,7]. Current forest canopy height
inversion methods based on PolInSAR include the ground phase difference method, the
coherence amplitude method, the combined-phase coherence amplitude inversion method,
and the two-layer random volume over ground (RVoG) scattering model. Among these,
the RVoG three-stage algorithm is the most commonly used and has been successfully
applied to various frequencies, including C-, L-, P-, and even X-band data [8–11], with
various forest types included in [12–14]. This method is based on interferometric complex coherence distribution characteristics, which are used to solve for the ground phase
and construct a lookup table (LUT) to invert the forest height by setting a reasonable
extinction coefficient and forest height threshold [15–17]. However, unbiased estimation of
the ground phase is impossible in RVoG models affected by temporal decorrelation and
variations in topography, vegetation, and baseline, as temporal decorrelation represents an
important factor affecting forest height estimation. Mette et al. [18] found that temporal
decorrelation leads to large errors in inversion results based on a study of three error
sources in vegetation height inversion using the RVoG model. Therefore, to improve the
inversion accuracy, it is necessary to reduce the errors caused by temporal decorrelation.
Lee et al. [19–23] studied temporal decorrelation using L- and P-band SAR data and found
that temporal decorrelation not only reduced the coherence coefficient but also increased
the volatility of the coherence phase in vegetated areas. Papathanassiou and Cloude [17]
proposed the RVoG + VDT and RMoG models; however, these approaches are limited by
an excessive number of model parameters, complex solution processes, and low efficiency,
which reduce their generalization. In addition, the ground-to-volume magnitude ratio
is usually assumed to be zero in the RVoG model; however, this assumption is not fully
valid in practice, especially in areas of low forest cover [20,21]. Lee [22,23] showed that
the temporal decorrelation effect is more severe in low vegetation areas; in these areas,
temporal decorrelation causes increased estimates of the volume coherence phase center
height, leading to overestimation of the low canopy, which is a common problem in forest
height estimation using PolInSAR.
In addition, the height estimation error caused by the penetration effect of microwave
signals in forests is commonly disregarded in forest height inversion by InSAR and PolInSAR. Previous studies have typically assumed that penetration of C- and X-bands into
the forest canopy is minimal; thus, InSAR and PolInSAR heights represent the true forest
canopy surface height [24–26]. However, larger penetrations have also been recorded in Xand C-bands [27,28];
for example, Kugler et al. [29] found that TanDEM-X data penetrated
up to 12 m in boreal and temperate forests, with significant differences in penetration
depth observed between the growing and defoliation seasons. C-band InSAR data obtained
from microwave remote sensing experiments in Indonesian tropical forests (INDREX 1996)
showed that the one-way extinction coefficient was 0.15–0.3 db/m, with a penetration
depth of 3–7 m, which was much less than the forest height [30]. TOPSAR results show
that the extinction coefficient was around 1 db/m, with a penetration depth of 4 m; for
boreal coniferous forests, the penetration depth is 11–22 m at an extinction coefficient of
0.2–0.4 db/m [15]. For L-band SAR data, stronger penetration can accurately reflect vertical
forest structure information, especially in tropical rainforests. However,
large penetration
Remote Sens. 2022, 14, 6145 3 of 27
depths cause a downward shift of the phase center, resulting in an underestimation of
forest height, an issue that has not been effectively resolved in previous studies. In response
to the forest canopy height estimation errors caused by microwave signal penetration
into the forest, Dall [31] the only theoretical framework published to date to estimate the
penetration depth and height bias of infinitely deep volumes; this represents an extremely
useful resource for correcting errors in the penetration of SAR data, and the theory was
validated by the results reported by Michael Schlund [10].
In summary, the effects of temporal decorrelation and penetration lead to tall canopy
underestimation and low canopy overestimation in the RVoG model of forest canopy height
inversion, representing an important error source in this model. To address these issues,
in this study, we used unmanned aerial vehicle synthetic aperture radar (UAVSAR) data
obtained from the AfirSAR project in 2016 as a basis to analyze the relationship between
microwave penetration depth and low canopy overestimation/tall canopy underestimation
in the RVoG model; we then corrected the forest canopy height estimation error of the
RVoG model using the penetration depth to improve the accuracy of forest height inversion.
The corrected results were validated using the RH100 LiDAR relative height variable. The
purpose of this research is to explore an error correction method for PolInSAR canopy
height estimation to serve global forest parameter estimation for spaceborne LiDAR (GEDI
and ICESat-2) in collaboration with spaceborne PolInSAR (e.g., ALOS-2 and the upcoming
TanDEM-L and BIOMASS satellites and NISAR programs).