Plant responses to biotic and abiotic legacies left in soil by preceding
plants is known as plant–soil feedback (PSF). PSF is an important mechanism
to explain plant community dynamics and plant performance in natural and
agricultural systems. However, most PSF studies are short-term and
small-scale due to practical constraints for field-scale quantification of
PSF effects, yet field experiments are warranted to assess actual PSF effects
under less controlled conditions. Here we used unmanned aerial vehicle
(UAV)-based optical sensors to test whether PSF effects on plant traits can
be quantified remotely. We established a randomized agro-ecological field
experiment in which six different cover crop species and species combinations
from three different plant families (
Plants influence biotic and abiotic soil properties and these changes can last in soil even after the plant is no longer there. This soil legacy of plants can feedback to the performance of subsequently grown plants such that their growth is enhanced or suppressed relative to growth in soil without a plant legacy (Brinkman et al., 2010). In recent years there has been a growing interest for understanding this mutual interaction process known as plant–soil feedback (PSF), because of its importance as a mechanism to understand plant community dynamics such as plant succession, exotic plant invasion and biodiversity–productivity relations (Kulmatiski et al., 2008; Bever et al., 2010; van der Putten et al., 2013). Also in agricultural systems, PSF is highly relevant and is one of the main reasons for practising crop rotations so that the risk for negative PSF can be kept low; however, mechanistic understanding of PSF is needed in order to make use of the potential of generating positive PSF effects (van der Putten et al., 2013; Dias et al., 2014). The vast majority of PSF studies has been conducted under highly controlled laboratory or greenhouse conditions and in order to adequately assess the impact of PSF in real ecosystems there is an urgent need to test PSF in the field and develop methodologies that facilitate this (Kulmatiski and Kardol, 2008; van der Putten et al., 2013). At the same time agronomic field studies are still vastly reliant on labour-intensive, time-consuming, destructive, and selective in situ data collection by experts (Nebiker et al., 2008). It is argued here that advancements in remote sensing platforms (i.e. Unmanned Aerial Vehicles) and imaging spectroscopy combined may offer novel opportunities for non-invasive assessment of plant trait responses to soil legacies at enhanced spatial-temporal resolutions (Faye et al., 2016) and unprecedented detail (Fiorani and Schurr, 2013), with potential use in both fundamental and applied research in natural and agro-ecosystems.
During the last decades, plant trait-based ecology has developed fast and is enabling a better mechanistic understanding of ecosystem processes across spatial and temporal scales (Cornelissen et al., 2003; Wright et al., 2004; Kattge et al., 2011; Díaz et al., 2016). Important plant traits from the perspective of ecosystem functioning comprise of physical and chemical plant characteristics such as plant stature and plant N content (Cornelissen et al., 2003; Díaz et al., 2016). Trait-based studies in plant ecology have mostly focused on natural ecosystems and their responses to natural and human-imposed disturbances (e.g. Garnier et al., 2007; de Bello et al., 2010). However, recently trait based approaches are being used to predict plant legacy effects in soil and subsequent plant responses (Orwin et al., 2010; Ke et al., 2015; Cortois et al., 2016). To date most PSF experiments have focused on plant biomass as sole measure of plant responses to soil legacies. However, there are a number of plant traits that are highly relevant for plant performance in both agricultural and natural systems as they represent aspects of plant quality (N content), competitive ability (plant height) and potential activity (chlorophyll content) which are also highly relevant for plant growth modelling.
Plant attributes invoke diverging interactions (i.e. absorption, reflection, and transmission) with light over different wavelengths (Pinter et al., 2003; Homolová et al., 2013). Consequently, spectral remote sensing has proven an effective source of information for monitoring vegetation in the field, non-invasively and comparatively efficiently, for diversified applications in past agronomic and ecological studies (Jones and Vaughan, 2010; Thenkabail et al., 2012), including species classification (Franklin, 2001), quantification of biophysical or biochemical plant constituents (Mulla, 2013; Qi et al., 2012), and multi-temporal monitoring of plant development (Zhang et al., 2003). Advancements in imaging spectroscopy are particularly relevant in this respect (Ortenberg, 2012; Fiorani and Schurr, 2013). Imaging spectrometers allow detection of subtle variations in spectral reflectance of the plant canopy by acquiring data in large numbers (up to hundreds) of contiguous narrow spectral bands (Campbell and Wynne, 2002; Warner et al., 2009; Qi et al., 2012). They invoke increased sensitivity to multiple crop traits (Homolová et al., 2013) and are therefore superior to multispectral alternatives (Shippert, 2004; Govender et al., 2007) regarding accurate discriminatory mapping and retrieval of vegetation traits (Rascher et al., 2011; Thenkabail et al., 2012; Kooistra et al., 2014). A variety of vegetation indices (VIs), embodying a mathematical manipulation of raw spectra from two or more wavelengths, have been conceived for vegetation monitoring purposes (Goswami et al., 2015) and were demonstrated to be stronger related to distinct plant traits than individual wavelengths due to isolation and enhancement of the spectral signal (Chuvieco, 2011).
It has been argued that conventional ground-based, airborne, or space-borne platforms are largely unable to provide remote sensing data at an adequate spatial (cm-level) and/or spectral resolution, repeatedly and at affordable costs for small-scale crop and vegetation field experiments with a large number of individual plots (Berni et al., 2009; Zhang and Kovacs, 2012; Colomina and Molinda, 2014). Unmanned aerial vehicles (UAVs), providing access to images with sufficiently high and flexible spatial-temporal resolutions at competitive costs and at an acceptable operational resilience, have received increased attention in related fields such as agriculture (Berni et al., 2009; Rango et al., 2009; Zhang and Kovacs, 2012; Honkavaara et al., 2013), and plant phenotyping (Chapman et al., 2014; Haghighattalab et al., 2016). Furthermore, proper plant trait retrieval methods and the associated accuracy (i.e. geometric and/or radiometric) and resolution(s) thereof require thorough evaluation (Lelong et al., 2008; Hardin and Jensen, 2011; Hruska et al., 2012).
The objective of the present study was to (i) develop and demonstrate a methodology for plant trait analyses using UAV-based imaging spectroscopy data, (ii) to assess the resultant accuracy for plant trait retrieval, and (iii) to evaluate the ability to discriminate plant trait responses to different plant legacies in soil in a field-based PSF experiment using UAV-based imaging spectroscopy data. We expected that UAV-based optical sensors can detect and quantify the plant traits (height, fresh biomass, N content, C content, and leaf chlorophyll content) at adequate resolution and accuracy. We also expected that plant trait responses to plant legacies quantified via in situ (i.e. ground-based) measurements can be assessed as well using UAV imaging spectroscopy analyses.
The investigation was conducted within a large-scale field experiment (Barel
et al., 2017), aimed at uncovering the influence of legacies of various major
crop species and combinations of cover crops on succeeding plants. A UAV
campaign and corresponding destructive sampling were conducted to retrieve
airborne imaging spectroscopy data and in situ oat (
The field experiment was established in spring 2014 (Barel et al., 2017)
to investigate the legacy of various species and species combinations of
cover crops on subsequently grown main crops. The study site (Fig. 1) is
located at the agricultural field facilities of Wageningen University &
Research (51
The experimental field as imaged from the Hyperspectral Data Cube
(HDC) acquired on 1 July 2015 represented as true colour RGB image, regions
of interest (ROIs) for
Schematic overview of individual plots and the approximated location at which samples for plant traits were collected.
Samples for plant trait analysis were acquired at the grain-filling stage of
the 2015 growing season in each plot cultivated with
Airborne imagery of the study site was acquired on 1 July 2015 on a cloud-free day by the Unmanned Aerial Remote Sensing Facility (UARSF) of Wageningen
University. The flight was conducted using an octocopter UAV (Aerialtronics
Altura AT8) carrying a custom-built Hyperspectral Mapping System (HYMSY)
sensing platform (Suomalainen et al., 2014), consisting of a pushbroom
spectrometer (Specim ImSpector V10
Next a crop surface model (CSM) was produced. Firstly, the DSM was derived from the RGB images using Agisoft PhotoScan Pro (v1.1.2) at a pixel size of 2.9 cm. Then, the areas between plots in the DSM were interpolated to retrieve an approximated ground surface digital elevation model (DEM) also in crop-covered areas. The DEM was subsequently differenced with the DSM to produce the CSM depicting within plot variation of estimated plant height (Fig. 3).
Visual three-dimensional representation of the crop surface model (CSM), upon differencing of the original digital surface model (DSM) and a secondary model approximating the ground surface digital elevation model (DEM) of the study area. Vegetation height is illustrated from low to high by change in colour from brown to green.
To extract imaging spectrometer and canopy height data for each experimental plot, region-of-interest (RoI) polygons were manually drawn for each plot. A 30 cm border was excluded from ROIs to retrieve average plot reflectance spectra from the HDC and height from the CSM while minimizing edge effects. Inspection of the RGB orthomosaic identified significant within-plot physical heterogeneity in 13 individual plots. We believe this was caused by accumulation of pathogens and/or nematodes under a distinct treatment. Due to the resultant conflict with the assumption of plot homogeneity required for the analysis (see Sect. 2.2), these plots were removed from the final analysis. Incorrect preprocessing of the data resulted in the cut-off of one additional plot, lowering the number of analysis objects to 41 monoculture and 15 bi-culture plots.
Overview of existing vegetation indices that were evaluated in this study for retrieving plant traits from optical remote sensing images. A division is made in two-band indices based on simple ratios and NDVI, and other indices using more than two bands. The index description includes their formulation and the original references.
The resulting dataset, consisting of 56 plots, was randomly split in a
calibration (50 %) and validation (50 %) set, provided that all cover
crop treatments were equally divided. For both sets, the Pearson
product-moment correlation coefficient was calculated to determine the
relations between the four selected plant traits, while also the correlation
of these traits with the height determined from the CSM was evaluated. Next,
the calibration set was used to establish relationships between the airborne
UAV data and in situ measured crop traits through (i) (univariate) linear
regression of a selection of existing vegetation indices (VIs) based on their
demonstrated success for correlating well with the traits presented here
(Table 1); (ii) derivation of alternative two-band VIs (Aasen et al., 2014);
and (iii) adopting full-spectrum partial least squares (PLS) regression. For
derivation of alternative VIs (ii), an optimization algorithm was written in
R to generate correlation matrices considering all possible (8836) band
combinations in simple ratio (SR), normalized difference (ND), and simple
difference (SD) vegetation indices (Aasen et al., 2014). PLS regression
followed earlier described procedures (Hansen and Schjoerring, 2003; Nguyen
and Lee, 2006; Cho et al., 2007; Abdi, 2010; Yu et al., 2014). The optimum
number of latent variables to include in PLS models was based on the minimum
predicted residual sum of squares (PRESS) during leave-one-out cross-validation (LOOCV), in agreement with Nguyen and Lee (2006). The performance
of calibrated models was assessed using the coefficient of determination
(
Summary of descriptive statistics for all plant traits measured in the field. LCC: leaf chlorophyll content; SD: standard deviation; CV: coefficient of variation.
The prediction ability of the best performing calibrated models per trait was
subsequently evaluated on the independent validation dataset. Prediction
precision and accuracy of the models was assessed by means of the coefficient
of determination (
The full-factorial randomized field experiment with different treatments of
preceding plant species resulted in various degrees of variation in the
traits of the subsequently grown test species
Spectral signatures for different plant treatments displayed deviations along
a vertical rather than a horizontal axis, i.e. the relative shape of
signatures was largely identical for all treatments (Fig. 4). In the visible
spectrum, the highest (6 %) and lowest (3 %) reflectance were
recorded at 555 and 675 nm, i.e. the chlorophyll absorption minimum and
maximum, respectively (Broge and Leblanc, 2000; Haboudane et al., 2002;
Vincini et al., 2007). Beyond the chlorophyll post-maxima (
Average reflectance spectrum of oat grown in the different
experimental plots and their associated treatments for the calibration (
CSM height was positively correlated to all crop traits, particularly for
validation plots (Table 3). In general, the observed interdependencies
confirmed the associated relationships between vegetation height and
variables such as growth rate, biomass, and plant fertility/health (e.g.
Cornelissen et al., 2003; Tilly et al., 2014). Strongest correlations were
observed for in situ measured crop height, indicated by correlation
coefficients of 0.85 and 0.91 for calibration and validation data,
respectively. Furthermore, relative variations in CSM height were also
significantly (
Correlation coefficients (
Coefficients of determination (
In situ measurements were linearly regressed with a selection of
well-established VIs (Table 1) based on the best matching bands from the HDC,
the main product of the hyperspectral mapping system. Regression analysis yielded
highly varying
Generated optimized indices for different plant traits and their wavelength dependency.
In agreement with the wavelength dependency of REP (670, 700, 740, 780 nm)
and MTCI (680, 710, 755 nm), the best performing two-band indices
recurrently exploit the near-infrared (
Statistical parameters of the PLS model calibration for the four selected plant traits. LCC: leaf chlorophyll content.
In order to explore the applicability of alternative band combinations, plant
traits were linearly regressed against all possible simple ratio (SR,
The hotspots identified for SRs largely aligned with those found for NDVIs,
and to a smaller degree with optimized SD indices (Table 5). In accordance
with the earlier findings for existing indices (Table 4), the best
performance was observed for indices borrowing from the red-edge
(
Finally, spectra were related to plant traits employing two partial least
squares (PLS) regression models. The first model (PLS1) incorporated all mean
plot reflectance measurements in the 450–915 nm range, the second (PLS2)
included plot-wise height measurements derived from the CSM as an additional
explanatory variable. The optimum number of latent variables (NLV) in the
PLS1 models ranged from 1 and 3 for fresh biomass and height to 5 and 11 for
LCC and N content, respectively. The NLV in PLS2 models for height and N
content changed to 5 and 2, respectively. The model precision and accuracy
was highest for height, LCC and N content, indicated by the
The factor loadings indicated the relative importance of explanatory
variables for the construction of each LV, i.e. higher loadings attribute
comparatively more influence (Hansen and Schjoerring, 2003; Nguyen and Lee,
2006). It was observed for all traits in PLS1 models that the first loading
weights allocate significant leverage to longer red-edge and near-infrared
wavelengths in particular. High loading weights for the second component were
recorded at 710 nm for LCC, and at 560 nm in the green peak for height and
N content. In all PLS2 models, CSM height was accredited with the highest
loading score for all traits. Consequently, the PLS2 model for fresh biomass
(NLV
Overview of validation statistics for the best selected existing and
new indices and both PLS models for the different plant traits,
LCC
The independent validation dataset was employed to assess plant trait
prediction accuracies of previously calibrated models, including the three
best performing existing indices, one of each optimized new index and the
best of two PLS models for each trait (Table 7). The highest prediction
accuracies were obtained for crop height (NRMSE
The plant legacies resulted in significant differences (
Mean and standard deviations of observed
Across all traits and analysis methodologies the red-edge and near-infrared
spectral region were consistently of critical importance, in contrast to
visual wavelengths. This finding is in agreement with expectations based on
univariate correlations of traits over wavelengths (not shown). The red-edge
slope is of particular relevance for leaf chlorophyll content (LCC) because
of its enhanced sensitivity to varied and higher chlorophyll levels while
circumventing saturation problems as observed in the blue and red due to
vast chlorophyll-induced absorption (Gitelson, 2012; Kooistra and Clevers,
2016). Reliance on wavelengths at the onset of the near-infrared follows
from the gradual stabilization of reflectance beyond the red-edge, which
settles at higher values for increased chlorophyll levels (Lamb et al.,
2002). Largely similar spectral regions were structurally highlighted for N
content, resulting from the inherent biochemical linkage between leaf N,
chlorophyll molecules and photosynthetic capacity (Sellers et al., 1992;
Weiss et al., 2001; Netto et al., 2005; Wu et al., 2008). Consequently,
wavelengths positioned in the red-edge were found to be highly sensitive for
chlorophyll absorption behaviour and thus to accumulation of nitrogen
(Thenkabail et al., 2012; Zhao et al., 2014). Although a direct physical
relationship between plant height and reflectance is absent, alternative
structural parameters (e.g. biomass and canopy densification) may serve as a
proxy for the former (Wang et al., 2011). Resultantly, the employing of
near-infrared wavelengths regarding plant height possibly followed from
It is well known that plant traits vary according to plant growth stage. As
plants mature and start senescing, stocks of both N and biomass are gradually
re-allocated to grains, hereby invoking reduced photosynthetic capacity,
discolouring of leaves and exposing of other plant pigments (Peinetti et al.,
2001; Murphy and Murray, 2003; Ciganda et al., 2009). Consequently, various
previous studies found that estimation of N (Zhao et al., 2014), biomass
(Yang and Miller, 1985) and height (Scotford and Miller, 2004) in mature
vegetation is prone to larger inaccuracies compared to in vegetation in
earlier growth phases. In our study measurements were obtained during
grain-filling stage, so in mature plants. The loss in photosynthetic capacity
in senescing plants likely partially explains the reduced importance of
visible (i.e. red) wavelengths, as chlorophyll absorption at these
wavelengths becomes less pronounced than generally is the case during
preceding stages (Gitelson, 2012). Besides, the maximization of biomass
accumulation in matured plants (Malhi et al., 2006) may explain lower
The different cover crop treatments resulted in marked differences in several
plant traits of the following crop of
We were able to pick up significant differences between the treatments both on the in situ measured and on the remote-sensing-based modelled values. These results provide scope for unmanned aerial vehicles (UAVs) and imaging spectroscopy as an enabling means to transfer PSF studies and related studies on legacy effects in soil to outdoor field environments (Fiorani and Schurr, 2013; Faye et al., 2016). To date most studies on plant legacies in soil and their impact on subsequent plant growth have been performed under controlled greenhouse conditions at small scale, yet it has been advocated that outdoor field experiments are needed to assess the magnitude and relevance of plant legacies (Kulmatiski et al., 2008; van der Putten et al., 2013). Here we propose, based on the results of our current study, that the use of UAV-based optical sensors allow for adequate field observations that will enable to complement and/or verify associate studies executed in controlled indoor environments. The use of UAVs is faster and more cost-efficient compared to conventional (i.e. hand-held) means, while limiting the intrusion of changing atmospheric conditions to affect measurements (Chapman et al., 2014). Moreover, UAVs enable operational resilience, besides adequate scaling of spatial detail and temporal revisiting times without the need for destructive sampling to measure and monitor ecological phenomena such as successional physiological vegetation processes over time (Faye et al., 2016).
Apart from differences in plant traits between plant physiological stages it also has to be noted that the quality of the predicted values is dependent on constraints invoked by the quality and quantity of ground truth data (Michaelsen et al., 1994). In situ sampling was conducted at diversified densities for different traits and/or for monoculture and biculture plots (see Sect. 2.2). Following from hypothesized plot trait and treatment homogeneity, samples were considered representing the remainder of plots. However, some degree of within plot heterogeneity was present. Consequently, calibration and validation of relationships between plot averaged spectra and field samples at one or a limited number of locations may have been suboptimal. To enhance the robustness of the models we therefore advise future studies to use a more extensive sampling layout such that field sample locations more accurately align with UAV spectrometer data from which data are further processed (von Bueren et al., 2015). Furthermore, improvements can be made by using a different flight, performed on a subsequent day, as a validation dataset to evaluate retrieval model sensitivity and by performing more flights over the growing season to capture temporal variation (Capolupo et al., 2015). The processing of the data to derive plant trait indices from the spectra collected using UAV-mounted sensors can be improved by making use of bootstrapping to find the best combinations of indices (Souza et al., 2010) and machine learning techniques based on the available spectral wavelengths (Singh et al., 2016).
Plant–soil feedback (PSF) studies gained scientific interest over the last decades, however field studies are urgently needed in order to evaluate the role of PSF processes under field conditions (Kulmatiski et al., 2008; van der Putten et al., 2013). Here we show that UAV-based hyperspectral remote sensing of plant traits enables to non-destructively quantify plant traits that respond to plant legacies in soil. This finding offers great potential to expand studies of PSF effects from the greenhouse to field settings. The plant traits that could be most accurately and precisely quantified were plant height and leaf chlorophyll content. The non-destructive nature of the measurements, after thorough parameterization, furthermore enables studying PSF effects at field scale at relevant spatial-temporal resolutions, this in turn will facilitate the elucidation of the underlying mechanisms.
The data will be made available via publicly accessible data repository
Dryad (
The authors declare that they have no conflict of interest.
We thank the staff from Unifarm and Irene Garcia Gonzalez and Dominika Piwcewicz for help with the field work. This work was supported by an NWO-ALW VIDI to GBDD (grant no. 864.11.003). Edited by: A. Rammig Reviewed by: M. Tuohy and one anonymous referee