1. Introduction
The capabilities of Synthetic Aperture Radar (SAR) data for forestry applications have been explored by a large number of studies. Several surveys employ the magnitude of repeat pass interferometric (InSAR) coherence |
γ| as the biomass estimator (e.g., [
1–
4]). The rationale for this method is that increasing growing stock volume (
GSV) typically results in increasing volume and temporal decorrelation and thus decreasing |
γ|.
In the boreal zone, the pronounced seasonality needs to be considered in the SAR data exploration. During winter the trees are commonly frozen, resulting in a deeper penetration of the incoming electromagnetic (EM) wave into the canopy volume [
5]. The backscatter generated by the trees as well as the contrast between forest and non-forest is reduced [
6–
8]. In the Siberian winter, the environmental conditions are stable. Due to the very low temperatures, the snow is dry and causes minimal scattering in L-band [
6,
9]. Since the soil is also frozen, changes in soil moisture do not occur. With regard to |
γ|, these conditions lead to very low temporal decorrelation for open areas. Even large temporal baselines of several weeks are not necessarily associated with large temporal decorrelation [
2,
4].
Several studies indicate that |
γ| images acquired during frozen conditions do indeed have potential for forest
GSV estimation [
2,
4,
10]. However, most studies aiming at
GSV retrieval from |
γ| data do not consider potential effects of differing tree species, although the shape of the tree crown is known to impact the location of the scattering phase center [
11–
13] and thus the magnitude of volume decorrelation. Also, the impact of forest type (plantation
vs. natural forest) and structure (tree height and density) on |
γ| has not yet been assessed [
14]. Preliminary investigations in Siberia using ERS-1/2 tandem data did not show a clear impact of tree species [
14]. On the other hand, [
14] achieved the most substantial |
γ|-
GSV correlation for an area dominated by larch.
There are a number of papers investigating the seasonal variability of C-band |
γ| over forest [
1,
15–
17]. Mostly, coherence was found to increase in winter, in particular under frozen conditions. This effect is higher for deciduous forests, which are defoliated during winter [
15]. Homogenous stands featuring a low proportion of deciduous species appear to have a higher temporal stability [
16]. During the growing season coniferous trees were found to feature higher coherence than deciduous trees [
1,
15]. Unfortunately, none of the studies distinguishing several species provides the
GSV distribution of the species studied. Other authors also working in boreal forests (Siberia) using ERS-1/2 tandem data [
18] observed a small impact of tree species on the |
γ|-
GSV relationship, but suggest further investigations are required. The study on hand follows this suggestion and addresses the impact of tree species on |
γ| using Phased Array type L-band SAR (PALSAR) data over Siberian forest at frozen and non-frozen conditions and emphasizes consequent implication for
GSV estimation.
4. Results and Discussion
The observed standard deviation of |
γ| over dense forest is higher under non-frozen conditions (see
Figure 2b), which is in accordance with former studies [
2,
4] based on Japanese Earth Resources Satellite (JERS-1) and PALSAR data. Furthermore, it was observed that the average of |
γ| of dense forest is significantly higher (confidence > 99.5%) under non-frozen conditions. The significance was evaluated applying the t-test resulting in a t-value of 3.3. This value is greater than the 0.995-quantile of the t-distribution:
t(0.995; 70) = 2.648. This issue of increased |
γ| during unfrozen conditions was thoroughly analyzed in [
4]. It was suggested that the deeper penetration of the incoming SAR signal results in increased volume decorrelation. Furthermore, the observations revealed increased temporal decorrelation over dense forest during frozen state.
Figure 3a–d provides a typical example of a PALSAR based
GSV-|
γ| scatterplot (do not consider the symbols for the species at this time). Under frozen conditions (
Figure 3a,b), an increment of
GSV results in decreasing |
γ|. Under unfrozen conditions (
Figure 3c,d), this trend is hardly visible. This is partly due to the spread of |
γ| at all
GSV levels. One potential reason for this spread is the spatiotemporal variability of environmental conditions during the growing season. Soil moisture and vegetation conditions can vary spatially resulting in variable temporal decorrelation for the same
GSV level. In winter, however, trees and soil are frozen, and little variation occurs in moisture. Precipitation in the form of very dry snow does not cause significant spatial differences of temporal decorrelation for L-band data [
6,
9,
26]. Another potential reason for the dissimilar observations between unfrozen and frozen conditions is related to forest structure and tree species. During unfrozen state, differing tree species and forest structure can be accompanied by varied temporal decorrelation. Geometric properties, such as crown shape or alignment of tree components, affect attenuation and the distribution of the major scatterers, and thus volumetric decorrelation [
11–
13].
During winter, under frozen and calm conditions, tree species-specific temporal decorrelation is unlikely. Geometric properties are also of less importance. Due to freezing the dielectric constant of the trees is reduced [
7,
27], resulting in decreased attenuation and a reduced InSAR phase height above the ground [
5]. Thus, the amount of scattering within the canopy is also decreased [
8] and the shape and structure of the canopy has less impact on the backscattered signal.
For the sake of clarity the data set shown in
Figure 3 was split into two groups. Under frozen conditions (
Figure 3a,b) all tree species follow the same trend and no obvious deviations occur. Under non-frozen conditions (
Figure 3c,d), differences between the species are apparent. For instance the coherence of birch and pine is much more scattered than that of spruce or aspen. Additionally, birch features by far the lowest minimum coherence. Nevertheless, from
Figure 3 no specific trend for the tree species can be observed. In
Figure 3 the overall spread of |
γ| is not caused by differences between species, but by the variations within the species. This observation, however, considers only one site and one interferogram.
Figure 4 summarizes species specific mean and standard deviations of |
γ| for dense forest based on all sites and interferograms. During non-frozen conditions all species feature higher coherence variability than under frozen conditions. Also, the species specific |
γ| are more diverse at unfrozen state (see also
Figure 5). The standard deviation of |
γ| is roughly 0.02 larger under non-frozen condition for all species. This finding is most likely caused by increased spatiotemporal variability of the environmental conditions during non-frozen conditions, as discussed above.
Figure 5 depicts the average deviation (difference) of the tree species specific |
γ| from the average |
γ| of dense forest for frozen and non-frozen conditions considering all sites and interferograms. Thus, it combines the information of
Figures 2a and
4a to summarize the observations. Under frozen conditions the deviation is rather small and reaches roughly +0.02 for larch and −0.03 for fir. Nevertheless, due to the low coherence level at frozen state this deviation is statistical significant (t-test, confidence > 95%). For the other species the mean deviations are much smaller. For example, the signatures of pine and birch do observably overlap. The confidence (t-test) that both species cause dissimilar |
γ| is < 60%. During non-frozen conditions, the deviation is increased and reaches values of up to +0.10 for fir and −0.05 for larch. In other words, during non-frozen conditions the |
γ| of fir is 0.15 greater than the |
γ| of larch, which is a significant difference (t-test, confidence > 99%). Thus, a clear impact on |
γ| is observable for some tree species under non-frozen conditions.
5. Conclusions
The impact of the tree species on |
γ| of dense Siberian forest was observed to be small in average, when frozen conditions are considered (in accordance to [
14] and [
18]). For non-frozen conditions the impact is increased against frozen conditions. Deciduous species (aspen, birch, larch) exhibit the lowest |
γ| (in accordance to [
15]). This impact of tree species on |
γ| under non-frozen conditions causes a fraction of the observed spread of the
GSV-coherence relationship (when the trend is investigated across all species). Thus, temporal and/or volumetric decorrelation to some extent depend on the characteristics of the specific species.
Moreover, increased intra-species variance of |γ| was observed for non-frozen conditions compared to frozen conditions. Assuming similar geometric properties of the trees of one species, the amount of geometric decorrelation should not differ much between the stands, as all stands feature dense forest. The large spread of coherence can rather be explained by varying temporal decorrelation caused by spatiotemporal variable environmental conditions as described above. Under frozen conditions, however, environmental conditions are much more stable in space and time resulting in decreased spread of |γ|. By all means, choosing SAR data acquired under frozen conditions is preferable when aiming at coherence based GSV assessment across species. When only data acquired under non-frozen conditions is available, the impact of the species on |γ| needs to be considered.