Nitrogen is a primary regulator of many leaf physiological processes, such as photosynthesis, respiration, and transpiration [1
], and is strongly linked to chlorophyll content, light use efficiency, and net primary production [4
]. Nitrogen is often a limiting factor for plant growth, and its role in the carbon cycle has been emphasized [7
]. Furthermore, nitrogen is an important input parameter in ecosystem process models [10
]. Acknowledging the significant role of leaf nitrogen in biodiversity and ecosystem functioning, leaf nitrogen content has also been proposed as one of the essential biodiversity variables by the remote-sensing and ecology communities for satellite monitoring of progress towards the Aichi Biodiversity Targets [12
Nitrogen has been retrieved with good accuracy using leaf- and canopy-level hyperspectral data despite the fact that it is only a relatively small constituent (0.2%–6.4%) in leaves [14
]. Hyperspectral data are capable of detecting the narrow absorption features of nitrogen by providing contiguous, narrow spectral band information. That offers an efficient and cost-effective solution to estimate leaf nitrogen compared to the traditional destructive sampling methods. Previous studies on nitrogen concentration estimation in vegetation used spectra from leaf powder, dry leaves, and fresh leaves, and also made estimates at the canopy level [16
]. There are a number of challenges in the retrieval of nitrogen at the canopy level, such as the mask of the strong water absorption [17
], the confounding effects arising from canopy structure, illumination/viewing geometry, and background [20
]. Efforts have been made to enhance the absorption features of nitrogen and reduce the sensitivity of the aforementioned parameters to canopy reflectance. Spectral transformation is one of the approaches, such as using first/second derivatives and log transformation of reflectance [22
]. Other approaches, such as continuum removal [17
], water removal [25
], and wavelet analysis [27
], also improved the nitrogen retrieval.
Leaf nitrogen has been determined in forest [18
], grassland [31
], and crop ecosystems [34
]. Empirical techniques are dominant in the retrieval of different vegetation variables from remote sensing data such as nitrogen, ranging from vegetation indices [37
] and traditional regression techniques such as stepwise multiple linear regression [17
] and partial least square regression [18
], to a number of artificial intelligence methods such as support vector regression, neural network, and Bayesian model averaging [32
]. Among the empirical techniques, vegetation indices are one of the simplest and most widely used approaches to estimate leaf biochemical contents such as nitrogen. Nitrogen mainly exists in proteins and chlorophylls in the leaf cells [15
]. Since nitrogen and chlorophyll are well correlated across different species [1
], vegetation indices designed for chlorophyll have been used as a means of nitrogen estimation [41
]. The spectral wavelengths near 550 nm and 700 nm as well as the red-edge region (680–780 nm) have been utilized for assessing chlorophyll [35
], resulting in a large number of indices [43
]. Compared with chlorophyll, there are a limited number of studies that propose indices specifically for nitrogen estimation, most of which were developed for crops [37
], with few developed for forest [49
Canopy structure confounds the estimation of foliar nitrogen when using canopy spectral data because it is the main driver of canopy reflectance variations. Ollinger et al.
] reported that the significant correlation between NIR reflectance (800–850 nm) and canopy foliar mass-based nitrogen concentration (%N) can be used for predicting nitrogen. However, Knyazikhin et al.
] pointed out that the relationship can be attributed to the correlation between NIR reflectance and canopy structure. Ollinger et al.
] argued that their hypotheses were based on the biological associations between nitrogen and structural traits that affect NIR scattering and reflectance. Additionally, Townsend et al.
] disagreed that the %N-NIR relationship is necessarily spurious, as Wright et al.
] and Ollinger [53
] indicated that the canopy structure and leaf properties may co-vary across plant functional types. Species and plant functional types (i.e.,
broadleaf and coniferous forest types) account for most of the variance of canopy chemistry that has been demonstrated across tropical [54
], temperate [58
], boreal forests [29
], and Mediterranean ecosystems [60
]. The link between species and canopy biochemistry can be explained by the concept of ‘global leaf economics spectrum’ [14
], which means that the key plant traits such as leaf mass per area, specific leaf area, leaf nitrogen, leaf phosphorous, leaf lifespan, and photosynthesis fall into a spectrum across plant species, and species converge towards the functional traits globally [14
]. However, the covariance of these functional traits has not been fully evaluated for nitrogen estimation.
There is a large body of literature focusing on the estimation of nitrogen in crops for monitoring and predicting crop yield [37
]. However, to our knowledge, little research has been conducted for validating the use of vegetation indices for mixed forests (including both broadleaf and needle leaf species), whose structure and composition varies substantially from that of crops. Such mixed forest is common in temperate zone at mid-latitude. This study aimed to evaluate the performance of 32 vegetation indices derived from airborne hyperspectral imagery for estimating canopy foliar nitrogen in a mixed temperate forest. The commonly-used partial least squares regression was performed for comparison. These vegetation indices can be classified into three categories that are mostly correlated to biochemical and physical properties of vegetation (i.e.
, nitrogen, chlorophyll, and structure properties such as leaf area index (LAI)). The nitrogen indices are chosen based on the effect of the physical basis of nitrogen absorption features on canopy reflectance. The chlorophyll and structural indices were involved here to exploit their potential for estimating nitrogen through the biological links between nitrogen, chlorophyll, and canopy structure.
The performance of three categories of vegetation indices related to the biochemical and physical properties of vegetation (i.e.
, nitrogen, structure, and chlorophyll) derived from airborne hyperspectral imagery was assessed to estimate canopy foliar nitrogen in a mixed temperate forest in this study. Partial least squares regression was performed for comparison. In this mixed temperate forest, functional type and species composition played the dominant role in explaining the variance of canopy foliar nitrogen. This is consistent with findings from different ecosystems, such as temperate, tropical, boreal, and Mediterranean ecosystems [29
Comparably accurate estimations of %N were observed across all three categories of vegetation indices. The best performing nitrogen-related indices utilized the physical basis of nitrogen absorption features in canopy reflectance, while the structural indices profited from the biologically functional links between nitrogen and canopy structure caused by functional type and species differences at the site. The best performing chlorophyll-related indices used the red-edge region and were subjected to the combined influences of strong chlorophyll absorption and structural properties. PLSR also captured the functional type variation, though it provided a lower estimation of %N compared with the best performing vegetation indices. The PLSR model obtained greater influence from the NIR and SWIR regions (Figure 5
). The lower accuracy of PLSR could probably be explained by the fact that only one latent factor was selected for the model (Section 3.4
). Due to the higher accuracy and ease of use of vegetation indices, canopy foliar %N was mapped by the best performing index, NDNI1510
The nitrogen related vegetation index NDNI1510
produced the most accurate estimation of canopy foliar nitrogen (R2CV
= 0.79, Table 4
) among all the vegetation indices involved in this study. Though NDNI1510
was derived in the heterogeneous Mediterranean shrub vegetation [49
], our study proved that it can also be applied to a mixed temperate forest. The index was developed based on the nitrogen absorption features at 1510 nm, with a causal basis, so in theory it should be generalizable to other species. The values of NDNI1510
paralleled those of canopy foliar nitrogen, and differed significantly between functional types (p
< 0.001). This indicated that functional types drove the relationship between the index NDNI1510
and nitrogen. We also examined other forms of NDNI that used nitrogen absorption bands other than 1510 nm, such as 1020 nm,1730 nm, 1980 nm, 2060 nm, 2130 nm, 2180 nm, 2240 nm, and 2300 nm [16
], which did not improve the estimation accuracy.
Other nitrogen-related vegetation indices derived for crops, such as NI_Tian [46
] and NI_Wang [47
], failed to estimate %N in this study site, indicating the poor transferability of such indices across vegetation types. The NI_Ferwerda developed in Ferwerda et al.
], based on multiple species using field canopy spectra, returned weak estimates of %N (R2CV
= 0.49, Table 4
). This can be justified by the differences between field and airborne canopy reflectance measurements; the latter were affected by the atmosphere, signal-to-noise ratio, etc.
Another possible reason is that the index could not take the composition and structure of the mixed forest into account. The index PALI proposed in Mobasheri and Rahimzadegan [76
] incorporated all nitrogen absorption bands; however, it failed to estimate %N when using the reflectance. When replacing the form of reflectance in the index with log(1/R) as suggested in Serrano et al.
], the modified PALI gave a good performance in estimating %N (R2CV
= 0.68, Table 4
). These findings were inconsistent with past studies that improved the accuracies in %N estimation using pseudo-absorbance [log(1/R)] rather than reflectance [22
The mean NIR reflectance between 800 and 850 nm provided an accurate estimation of nitrogen (R2CV
= 0.73, Table 4
), which was in agreement with the findings reported for some temperate and boreal forests in North America [7
]. In addition to NIR reflectance, most of the other structural related indices were capable of estimating foliar nitrogen at a moderate to good accuracy (Table 4
). The results might be explained by the control of functional type and species composition with regard to the ‘global leaf economics spectrum’ [14
]. That is the functional convergence across species among optically important leaf traits such as leaf mass per area, nitrogen concentration, and canopy structural properties such as LAI [53
]. However, a weak correlation between %N and LAI was observed in this study (R2
= 0.38, results not shown). A structural parameter, crown closure, was found to determine the relationship between nitrogen concentration and hyperspectral data in a boreal forest study [29
], since crown closure represents the amount of green foliage in the canopy and thus controls the spectral response of the canopy. However, that relationship was not observed in this study (R2
= 0.09, results not shown). The reason might be the relatively dense forest with a small range of crown closure (mean ± std: 0.82 ± 0.06) at our study site. The %N—NIR relationship displays a gradient across functional types; therefore, a broader context of canopy structure including foliage distribution, inner crown structure, and outer canopy surface should be considered to fully understand the mechanisms of the relationship.
First of all, the foliage clumping of needles in shoots rises the multiple photon-needle interactions within a shoot, thereby increasing the probability of self-absorption and reduced reflection [53
]. Secondly, additional factors may include crown shape, canopy volume and density, and gap fraction. [53
]. For instance, broadleaf stands have spherical or ellipsoidal shaped crowns and needle-leaf stands have ellipsoidal or conical shapes. Last but not least, the smooth and continuous upper surfaces of broadleaf canopies allow more photons to be detected by a sensor, which leads to higher reflectance [50
]. These factors, as a whole, result in higher reflectance of broadleaf as compared to needle-leaf trees across the NIR region and SWIR spectral regions (as presented in Figure 7
). Therefore, the functional link between %N and NIR reflectance could provide a means for estimating canopy foliar %N [51
] and a simple and rapid means of generating regional maps of nitrogen variations.
The best performing chlorophyll-related vegetation indices utilized the spectral information in the red-edge region, such as Boochs2, DDn, and Sum_Dr1, which gave accurate estimation of canopy foliar nitrogen with R2CV
of 0.76, 0.72, and 0.73, respectively (Table 4
). The red-edge region (680–780 nm), which has a low reflectance in red due to chlorophyll absorption and high reflectance in NIR due to leaf internal scattering and canopy structure, has been shown to be more sensitive to chlorophyll and nitrogen [35
]. Given the close relationship between nitrogen and chlorophyll as well as the solid background of remote detection of chlorophyll, the chlorophyll-related indices provide an indirect means of nitrogen estimation when hyperspectral data are not available for certain nitrogen indices. Sentinel-2 and -3 offer the potential of using red-edge chlorophyll-related indices for indirect nitrogen estimation in forest, which has been shown to be feasible in different crop species and grassland [99
]. However, the correlation between nitrogen and chlorophyll may become less strong in nitrogen-rich ecosystems [28
]; in such cases, care should be taken when applying the chlorophyll-related vegetation indices to nitrogen estimation.