Biodiversity describes the genetic and functional variability within and between all taxonomic ranks of life as well as within and between their organizational aggregates, such as communities or ecosystems. Biodiversity is a key element in the provision of ecosystem services influencing ecosystem functioning and stability [1–4]. Owing to land-use changes, climate change, invasive species and others [5], global biodiversity is in decline and predicted to remain so within the near future, although there is a high degree of uncertainty and future predictions of biodiversity trajectories vary by orders of magnitude [6]. Despite considerable political efforts up to the highest international levels, such as the Convention on Biological Diversity (CBD) [7] and its consecutive protocols, the estimated biodiversity loss is still unhalted [8].

One reason for the uncertainty in biodiversity estimates is the inherent difficulty in measuring and monitoring biodiversity. Nevertheless, in order to slow down negative biodiversity trends, knowledge about the spatial distribution and temporal dynamics of biodiversity on scales relevant to the conservation decision makers is needed. Indeed the CBD calls upon each party to monitor “through sampling and other techniques, the components of biological diversity [. . . ] paying particular attention to those requiring urgent conservation measures” and further to “identify processes and categories of activities which have or are likely to have significant adverse impacts on the conservation and sustainable use of biological diversity, and monitor their effects” [7].

Conservation planning is usually made on landscape to regional scales [9] where local plot information is too sparse and continental or global estimates are too coarse, yet the area to be covered is too large for direct inquiry. Remote sensing techniques promise to close this gap by providing spatially continuous information that can be related to point estimates of biodiversity at comparatively little cost. Moreover, remote sensing allows frequently repeated recording of environmental information, thereby providing time-series which are essential to monitoring and understanding biodiversity gradients and their temporal trajectories.

According to Turner et al. [10] two basic approaches in assessing biodiversity aspects from remote sensing data can be distinguished, namely direct and indirect ones. The former aims to detect the variable of interest, e.g., species or communities, by directly relating their occurrence to remote sensing derived measures, such as species-specific spectral signatures determined by leaf chemical constituents [11,12] or plant species cover based on species-specific phenology [13]. Indirect remote sensing of biodiversity refers to models based on remote sensing derived environmental variables such as primary productivity [14], time series of vegetation indices [15,16] or habitat structure [9,17].

In Europe, forest biodiversity is one of the major conservation targets and at the same time provides one of the major challenges for the establishment of conservation strategies at regional scales. This is because most of Europe’s typical natural land-cover would be forests, although much has been lost or converted due to human activities [18]. Using remote sensing technologies for direct mapping of plant biodiversity has been reported to work with promising results in open type ecosystems (e.g., [15,19–21]). However, forest biodiversity is difficult to assess due to their explicit three-dimensional structure. Passive remote sensing systems are only able to detect the forest surface, the canopy and the contributing tree species. Consequently many studies on forest plant diversity focus on trees (e.g., [12,22–27]). However, considering vascular plants, the most species rich element of temperate European forests is often the herb layer [28], while species richness of trees is mostly small.

Promising candidates for biodiversity mapping are hyperspectral and LiDAR remote sensing [10,29–31]. Hyperspectral data has been successfully used to model invasive plant species [32,33], tree species abundance [34], plant functional strategy types [35], plant species richness [21,23,36] and plant community gradients [37] due to very distinct variability in absorption and reflectance properties.

It is commonly expected that very high spatial resolution optical imagery (hyperspatial) is needed for remote sensing based biodiversity assessments [38], since hyperspatial imagery reflects individual small-scale structures such as tree crowns. However, increasing the spatial resolution brings with it also an increase in variability in the reflected signal, which complicates any classification task [26,39]. Hyperspectral remote sensing on the other hand allows to very precisely distinguish surface characteristics by recording additive sub-pixel spectral signatures [32], thereby reducing the need for extremely high spatial resolutions and avoiding the corresponding additional noise.

The key open question is whether canopy species composition can be a surrogate for sub-canopy communities. Assuming that communities are entities across forest strata, hyperspectral remote sensing can then be expected to allow modelling complete and sub-canopy forest communities by assessing only a fraction thereof, i.e., canopy composition. This of course will very much depend on the ecosystem and communities in question. However, independent of community relationships, another mechanism may come to aid: Hyperspectral data reflects environmental conditions acting upon plants, such as soil pH, water availability, nitrogen availability and others [40,41], which are known to influence species distributions and community composition. Low canopy nitrogen levels, for example, might then reflect substrate conditions which affect and shape herb layer communities.

On the other hand, active systems like RaDAR and LiDAR directly provide information on the sub-canopy layers of forests by penetrating canopies and preserving the position of reflecting objects. The structural information provided has been shown to be highly valuable for biodiversity modelling encompassing vegetation density distribution in different forest strata, which is an essential variable describing vegetation itself [42] but also for explaining distributions of other groups such as birds [43–45], forest beetles [9] or spiders [17].

It is expected that hyperspectral and LiDAR data will complement each other and result in improved modelling capabilities once combined [46], especially for modelling of multi-layered forests. While hyperspectral data provides detailed surface reflectance characteristics, LiDAR adds detailed three dimensional positional information. Studies using hyperspectral and LiDAR data for modelling biodiversity mostly report increased accuracy when using both hyperspectral and LiDAR derived information compared with when using only one of them, e.g., in forest tree species identification [25, 46,47].

These hyperspectral and LiDAR approaches have to be seen in context of upcoming space-borne sensors. While studies relying on airborne remote sensing are often limited in their spatial extent, civil hyperspectral space-borne sensors are about to become increasingly available, e.g., EnMap [48] or HySpiri [49]. Moreover, they will be complemented by further space-borne LiDAR and RaDAR systems, which may allow assessing three-dimensional vegetation structure from space [30].

The purpose of this study is to investigate which information on vascular plant diversity can be derived from remotely sensed data. In particular, the capabilities of hyperspectral and LiDAR predictors are analysed for different vegetation strata in complex forest ecosystems. The following two hypothesis were tested. First, plant species richness and communities of different forest layers can be modelled relying only on remote sensing data. The combination of LiDAR and hyperspectral predictors is expected to improve the predictive power of the models because they reflect complementary information. Furthermore, we hypothesised that LiDAR performs better than hyperspectral data in modelling sub-canopy patterns. We expected this because the former provides direct information on the vegetation while the latter can do so only via proxy relationships of canopy composition and biophysical state to sub-canopy communities and environmental condition. Second, we expected that reducing the dimensionality of the learning problem prior to modelling the biodiversity metrics would improve model predictions. We used up to 135 predictors, many of which are noise variables. Therefore, we expected an increase in model performance when only relevant predictors are used as input variables.

With this study we aimed to estimate the added value of combining hyperspectral and LiDAR remote sensing data for modelling vascular plant diversity patterns in complex forest ecosystems. We expected that, since both systems provide distinct information, they would complement each other, leading to an improved set of model predictors. However, we found that the added value of hyperspectral plus LiDAR data was small. While both systems achieved fair cross-validated accuracies in modelling vascular plant species richness and performed well in modelling aspects of overall, herb layer and tree layer communities, combining them did not result in improved models. This is unexpected, since the limited correlation between LiD and MNF variables suggested that each one contains distinct information (Figure 4).

Overall, LiDAR based models performed best with respect to all vegetation layers. Hyperspectral data only provided valuable information on species richness within the tree layer. This result was expected since the trees shape most of the spectral reflectance recorded by HyMap. Therefore, we conclude that the tree layer composition does not provide sufficient proxy information on the underlying strata concerning SR in our study area. In general, modelling species richness directly from remote sensing data is complicated by the fact that very different communities with different reflective properties may share the same species richness. A way to circumvent this problem is by exploiting relationships between variability in spectral reflectance and species diversity [23,36,81]. Such approaches, however, are limited to mapping of “optically accessible” species, i.e., either to studies in open systems or studies concentrating on canopy species diversity. In temperate forests, where the herb layer is the most numerous element of species richness [28], this approach can only work if species richness of the herb layer is correlated to tree layer richness. Indeed, Vockenhuber et al. [82] reported such a relationship in German lowland beech forests. The presented data in contrast did not share this relationship (τ = 0.05), rendering this approach unsuited to the Bavarian Forest study site.

Model results for the abundance-based Shannon index were poor in most cases, except for the shrub layer. This is in contrast to Oldeland et al. [21] who found that using abundance based estimators of biodiversity could increase model performance as compared to using species richness alone, albeit only in some cases and for open savanna vegetation. These differing results may be due to the fact that in complex forest communities, reflectance is only a function of a fraction of the occurring species, i.e., mostly canopy species, and fractional cover does not scale with abundance in the remotely retrieved image. Nevertheless, along those lines it is surprising that hyperspectral data could not explain the tree layer Shannon index.

Community composition based on NMDS ordination scores was modelled with great success, especially for overall and herb layer composition using either LiD or MNF predictor sets, whereas tree layer community composition was predicted better using MNF predictors and best by MNF+LiD. In the case of ALL and HL, the first NMDS axis mainly reflected species turnover along a canopy openness gradient, spanned by shadow-adapted, forest-dwelling species to light-demanding, non-forest species. This explains the exceptionally good models, as canopy density is a fundamental quantity measured by LiDAR [45,83]. With hyperspectral data this is less intuitive, although it has been used with moderate success in modelling forest structural attributes such as stem density [84]. Our results suggest that unlike for species richness, there is a linkage between canopy and sub-canopy communities, which explains the good performance of hyperspectral predictors.

Models predicting the second NMDS axis did not reach the performance of NMDS₁ models. Nevertheless, ALL, HL and TL, unlike SL, still achieved moderate performance. This can be attributed to the second ordination axis explaining less variability of the vegetation composition in general and additionally because it reflected species richness, which in itself could be modelled with only moderate success. This is in contrast to results from non-forest ecosystems. Schmidtlein et al. [85] successfully modelled three dimensions of NMDS scores of non-forested wetland vegetation with R CV 2 of 0.92, 0.82 and 0.78 using partial least squares regression based only on hyperspectral data, however with the precondition of homogeneous plots. Using the same approach, Feilhauer et al. [37] modelled open heath land-vegetation with R CV 2 of 0.74, 0.58 and 0.54.

The stagnation or even decrease in performance of several MNF+LiD models as compared to LiD alone could be due to the lower probability of LiD predictors to contribute to each tree and hence a methodological limitation. At maximum, 125 MNF variables were used together with 10 LiD variables, out of which only mtry predictors are sampled for node splitting. If there was such an effect, one would expect it to be reduced after feature selection, which reduced the numbers of MNF variables greatly, sometimes to equal numbers of LiD and MNF predictors. This, however, was not the case. Nevertheless, effects might be obscured by over-fitting as outlined below.

No studies have been published comparing LiDAR and hyperspectral data for modelling of plant species richness or communities. However, several studies have tested both for species classification. The results of this study regarding the added value of hyperspectral plus LiDAR data differ from those published in literature, where mostly improvements in classification accuracy are reported. Dalponte et al. [46] modelled 18 forest tree species using hyperspectral and LiDAR data and arrived at class [producer] accuracies of 40% to 90% with an increased accuracy for some tree species of up to 13% comparing only hyperspectral to the combination of hyperspectral and LiDAR data. Mundt et al. [86] reported an increase in overall mapping accuracies from 75% to 89% in mapping a single shrub species in a steppe environment when combining hyperspectral and LiDAR as opposed to using hyperspectral data alone. However, some studies also found that LiDAR provides improved models only in specific cases. Jones et al. [25] reported no overall improvements in classifying forest tree species, but specific improvements for some species by 5% to 12% in producer’s accuracy. Moreover, Onojeghuo and Blackburn [87] found a negative impact of including LiDAR derived canopy height and surface models in a maximum likelihood classifier for mapping a single vegetation class. However, using LiDAR for post-classification correction resulted in increased accuracies.

It is a challenge of increasing importance to correlate biodiversity on the ground with remote sensing data. As a matter of fact, the detection of plant species α-diversity from remote sensing imagery, such as species richness or the Shannon index, is difficult to achieve due to complex vegetation structures (canopies) and different plant sizes. However, there is growing awareness that at the relevant scale of ecosystems and communities, other aspects of biodiversity such as β-diversity are more important for detecting and explaining changes [88]. In detail there are various perceptions of β-diversity (see, [89,90]) but generally this aspect of biodiversity refers to qualitative differences between reference plots such as pixels in remote sensing data. Analysing spatial patterns of compositional contrast and integrating them to multi-plot comparisons between neighbouring datasets (e.g., raster cells) [91,92] allows to make a direct connection between the biotic information gathered on the ground and the remote sensing regarding spatial trends of dissimilarity and heterogeneity. When calculating values of dissimilarity, spatial grain (plot/pixel size) and distance decay must be considered [93,94]. Summing up, β-diversity measures derived from community composition can be seen as a promising tool for the future development of remote sensing of biodiversity, as they are not bound to individual species traits or numbers.

Testing our second hypothesis, we examined whether the high dimensionality of the input predictor sets had a negative impact on model quality. Our results clearly show that using only a reduced set of relevant predictors did not substantially benefit the predictive performance of the random forest models in most cases. This confirms claims of random forests being able to deal with large amounts of non-informative predictors, rendering any attempts for preceding feature selection unnecessary [95]. On the other hand, the step of selecting all relevant features [78] by means of the Boruta algorithm may not have been enough. Subsequent stepwise model building based on subsets of these features sensu Genuer et al. [72], i.e., the minimal-optimal feature set, may have improved the predictive performance, as the latter is often a unimodal function of the number of predictors in statistical models [69,74]. Nevertheless, given the computational cost of up to 10,000 fitted random forests per mtry_i, response and predictor variable in the presented cross-validation scheme, feature selection for random forests can be seen as a dead end.

There was substantial collinearity between the LiDAR predictors, which is a serious issue in statistical modelling. Random forests have been shown to be sensitive to collinearity among predictors concerning variable importance measures [96]. However, with respect to predictive performance, which was the target quantity of our analysis, collinearity can be assumed to be negligible, as long as the collinearity among predictors remains constant [97]. Nevertheless, extrapolating beyond the geographical or environmental range of training data may be problematic [98].

At a methodological level, this study calls for caution in relying only on OOB estimates for assessing prediction performance. The OOB accuracy estimates overlap very well with those obtained in ten-fold cross-validation unless feature selection is conducted in the presence of many noise variables. The increased difference between R OOB 2 and R CV 2 after feature selection may be due to over-fitting to the training dataset, although this is not expected to occur in random forests [65]. This is noted also by Dudoit et al. [99] who state that “the out-of-bag method [. . . ] for estimating misclassification rates results in overly optimistic estimates of the error rates” when preceded by feature selection.

So far, although we did not model species directly, species richness and communities were modelled by a direct approach sensu Turner et al. [10], i.e., we directly related reflectance to species communities or species richness, rather than first estimating environmental parameters (primary productivity, water stress, etc.) and then using these to model biodiversity variables. A promising next step is now to make full use of the information provided by the sensor systems and combining both direct and indirect approaches. This could mean including abiotic geomorphological predictors (slope, aspect, topographic heterogeneity, etc.) derived from the LiDAR terrain model, which have been shown to influence species distribution patterns at various scales (e.g., [100,101]). Moreover, these geomorphological variables, sometimes in conjunction with climate data, allow to model further predictors such as irradiation or soil wetness [42,102].

The sometimes limited predictive performance of the models might have been caused in part by the high variability in land cover types, which ranged from closed canopy forests to early-successional open areas resulting from bark beetle infestations and wind-throw. One does expect that decision tree based methods are flexible enough to deal with such difficulties. Nevertheless it remains to be tested whether, for example, a stratification of the response variable based on openness or “optical accessibility” can improve the models.

In order to model biodiversity, we broke down the problem to modelling indices and ordination scores separately for each axes. This discontinuous statistical analysis is one major limitation of the approach taken in this study, since the step of information condensation is accompanied by a loss of valuable ecological information such as taxonomic identity or multidimensional position in ordination space. We believe that further steps towards truly multivariate learning techniques are highly promising. A candidate method for direct prediction of communities, without the step of ordination or index calculation, might be for example multivariate random forests [103], yet these are still at the very early stages of development.

This study has analysed the potential of LiDAR (LiD) and hyperspectral (MNF) remote sensing for modelling vascular plant diversity patterns in complex forest ecosystems. Both are promising techniques for regional scale biodiversity modelling, but the information gain of their combination should be thoroughly evaluated. Overall, the added value of hyperspectral plus LiDAR data was found to be small, with a maximal improvement in cross-validated R CV 2 ≤ 0.07 across all analyses in this study.

For modelling overall species richness, LiDAR predictors were the best choice ( R LiD 2 = 0.3, R MNF 2 = 0.08, R MNF + LiD 2 = 0.2). Broken down into forest layers, LiDAR performed best for herb and shrub layer ( R LiD 2 = 0.26and 0.55, R MNF 2 = 0.03 and 0.02, R MNF + LiD 2 = 0.16 and 0.43, respectively), while hyperspectral predictors were better in explaining tree layer species richness, only surpassed by the combination of both ( R LiD 2 = 0.11, R MNF 2 = 0.17, R MNF + LiD 2 = 0.22). When community composition was of interest, both systems provided similar performance overall and in the case of the herb layer ( R LiD 2 = 0.75and 0.55, R MNF 2 = 0.76 and 0.49, R MNF + LiD 2 = 0.78 and 0.52, respectively). For the tree layer, predictors derived from hyperspectral data performed better than LiDAR predictors ( R LiD 2 = 0.28, R MNF 2 = 0.37, R MNF + LiD 2 = 0.39).

Despite the importance for European conservation strategies, modelling sub-canopy biodiversity patterns of temperate forests from remote sensing information has been rarely attempted so far and remains challenging. Our study aimed to exploit direct measurements of vegetation cover and indirect relationships between canopy reflectance and sub-canopy characteristics using machine learning algorithms and a rigorous cross-validation scheme. This approach is very promising, but ecological knowledge needs to be integrated further into remote sensing approaches. In particular, future research should focus on an integrated statistical analysis, which directly relates species or community information to remote sensing data, without the stepping stone of biodiversity indices or ordination scores.