We mixed a soil substratum consisting of sand, peat and meadow soil in equal parts. These components were homogeneously mixed and then sterilized overnight using a Mafac Sterilo soil sterilizer. The soil was put into pots of 40 cm diameter in several layers, each with long-term NPK fertilizer and with wicks for irrigation. Each pot was put on a bucket filled with water, with the bottom of the pot 1.5 cm above the water level. The five wicks dip into the water reservoir for supplying the pots with capillary rising moisture through the wicks.

Three four-year old seedlings (ca. 1 m in height) of European Beech (Fagus sylvatica L.) were planted into each pot before leaf flushing in spring. The trees were numbered and then randomly allocated to pots; subsequently, the pots were randomly allocated to dry group and control group. In total, 66 beeches were planted in 22 pots. Five of the pots were excluded from further analyses because the trees did not grow green leaves. Nine control pots and eight dry-group pots were included in the experiment. Additionally, the number of buds was counted on each tree in order to estimate the expected number of leaves. The young beeches were kept under a semi-transparent roof to shield them from rain and direct sunlight.

After an initial growth phase from March to August, half of the pots were cut off from water supply in mid-August, while the other pots (the control group) were continuously watered. The dry group was kept from water supply for five weeks until most signs of dryness stress were clearly visible and many leaves had become brown or rolled up.

Soil water content was measured daily using a TDR probe from the beginning of the drying experiment. Unfortunately, the TDR probe was available only for a limited period until mid-September. Leaf water content was measured at six points in time on up to 49 leaves from trees randomly selected for the leaf sampling. The fresh leaves were weighed, scanned to measure their area, then dried at 105 °C for 24 h and weighed again. Equivalent water thickness (EWT, g/cm², [18]) was determined by dividing the difference between fresh and dry leaf mass by leaf area.

Furthermore, leaf chlorophyll content was measured on several leaves of randomly selected trees using a Minolta Spad-502 chlorophyll meter [19,20]. The mean value of three measurements per leaf was recorded.

A hyperspectral scanner (NEO HySpex VNIR-1600, [21]) was mounted on a platform about 4 m above ground (Figure 1). The scanner was fixed on a rotation stage that moves the line of sight across the targets acquiring an image composed of scan lines. The instrument has an across-track field of view of 17°, resulting in lines of about 1 m width. The images covered an area of about 3 m × 1 m each. Each image contains two or three pots with near-nadir viewing. The scanner was equipped with a 3 m focal lens with about 0.5 m depth of focus so that most of the area was imaged in focus. A white reference target (Spectralon) of known reflectance was included in the scans in order to convert the recorded radiances to reflectance values. Black foam rubber mats were placed on the ground to reduce stray light from the background. The images were recorded under natural light conditions. Before each scan the integration time was adapted to current brightness. At each of the four imaging days, six images were recorded to cover the 17 pots considered, so a total of 24 hyperspectral images has been utilized in this study.

The camera uses a 2-dimensional CCD sensor array of 320 × 1,600 silicone detectors. One dimension is used for spectral separation and the other dimension is used for imaging in one spatial direction. The second spatial dimension is covered by movement of the sensor. The camera records 1,600 pixels across track, resulting in sub-millimeter spatial resolution. The pixel instantaneous field of view is 0.18 mrad across-track and 0.36 mrad along-track. The speed of the rotation stage was adapted so that an image with approximately square pixels is formed out of the single lines the camera recorded. By binning the 320 sensor pixels in spectral direction, 160 bands were recorded in the spectral range of 410 to 990 nm with a spectral sampling distance of 3.7 nm. Data was recorded in 12 bit radiometric resolution.

Image preprocessing consisted of sensor calibration (converting raw data to radiance and adjusting sensor sensitivity across the line of sight), spatial resampling (merging four pixels into one), division of radiance spectra by mean Spectralon radiance spectrum (conversion from radiance to relative reflectance) and finally spectral resampling for a better signal-to-noise ratio. In the spectral resampling step, the uniform bandwidths were transformed into bandwidths adapted to the signal-to-noise ratio in the respective wavelength region, resulting in 85 bands instead of the original 160 bands. The pixel size depends on distance to the sensor. At 3 m distance, the pixel size after spatial resampling is about 1 mm².

Figure 2 shows a time series of images in true color (RGB = 608 nm, 557 nm, 451 nm). The three pots and the Spectralon white reference can easily be spotted. Each pot contains three trees, but these cannot be discriminated from above. For this reason, spectra have only been extracted for pots, not for single trees. Mean spectra for the respective pots were extracted from the images by manually identifying image regions dominated by leaves. For each region of interest and each time step, the mean spectrum was saved (Figures 3 and 4). These spectra (one spectrum per pot per recording date) were used as training spectra in the following regression analyses.

Leaf reflectance in the 350–2,500 nm spectral range was measured at regular intervals using an ASD FieldSpec2 spectroradiometer with a Plant Probe (ASD Inc., Boulder, CO, USA). The Plant Probe consists of a Contact Probe, a spectrometer attachment with a light source covering a 10 mm radius, and a Leaf Clip, a fixture to clamp leaves between the light source/sensor head and an interchangeable black or white background. The Leaf Clip/Contact Probe assembly guarantees constant illumination and viewing geometry and shields the sample from ambient light [22,23].

A Prospect-5b [24,25] inversion was used as an independent method to estimate EWT and chlorophyll content from the resulting leaf reflectance spectra. We used the Matlab inversion routine by Stéphane Jacquemoud that can be downloaded at [26].

Spectral indices, like Water Index (WI) [27], Normalized Difference Water Index [28] or Simple Ratio Water Index [29], have been used to estimate vegetation water content from NIR spectroscopy data (refer to [30] for further examples). Among these, only WI, the ratio of reflectance at 900 nm and 970 nm, lies within our sensor’s spectral range. WI was able to explain only 0.36 of the variance of EWT measured.

In order to explore indices of all possible combinations of HySpex bands, we calculated normalized difference indices according to equation (1) (1) Index = Band 1 − Band 2 Band 1 + Band 2and determined correlation between the index values and measured EWT values. To mitigate overfitting, a leave-on-out cross-validation was applied. As can be seen in the 2D-correlogram in Figure 5 the indices were able to explain up to 0.79 of the EWT variance. The regions that correspond to the highest correlations are highlighted in red and orange in the correlogram [31]: The highest correlations exist between EWT and indices formed solely of bands in the blue spectral region around 450 nm. Further hotspots of correlation are located in the red and green regions (indices similar to PRI [32,33]) and in the region of indices, with one band in the red edge around 700 nm and the other band on the near infrared plateau. For other index types (ratio or difference indices), the results are similar.

Since high R² estimations of EWT based on two-band indices were spread across the whole spectral range considered, we decided to employ a regression technique that makes use of the full hyperspectral information. Thus, EWT was estimated with a partial least squares regression (PLSR, [34,35]). PLSR is an established robust tool that projects the data into a low-dimensional space formed by a set of orthogonal latent variables by a simultaneous decomposition of X (spectral matrix) and Y (chemical concentration matrix) that maximizes the covariance between X and Y[36]. The method is well suited for calibration on a small number of samples with experimental noise in both chemical and spectral data [37,38]. PLSR is comparable to principal component regression, but differs in that both matrices are simultaneously decomposed.

The PLSR was trained on spectra of the pots and the according gravimetric equivalent water thickness measurements. This leap of scales was intended as a test of PLSR’s ability to estimate a variable of interest from canopy scale spectra, but trained with leaf spectra. One PLSR was used for all time steps. After testing models with 1 to 20 latent variables, we selected a model with five latent variables. Prior to the leaf water estimation, the images were classified using an unsupervised k-mean classifier in order to create a mask of non-leaf pixels. Leaf water content was estimated only for leaf pixels using the PLSR coefficients on image spectra.

Because chlorophyll is optically active in the spectral region considered with main absorption bands in the blue and red regions, we decided to use a very simple and parsimonious method for chlorophyll content estimation. We calculated a linear regression between a simple vegetation index (Simple Ratio: ρNIR/ρRED, with wavelengths 790 and 670 nm) and Spad chlorophyll measurements for image-based chlorophyll content estimation.

Maps of chlorophyll content were created using this regression to transform Simple Ratio values to Spad values and the empirical polynomial transformation proposed by Markwell et al.[19] for transforming Spad values to chlorophyll concentrations: (2) y = 10.6 + 7.39 x + 0.144 x 2 ,where y is the chlorophyll content in μmol·m⁻² and x is the dimensionless Spad meter value. This equation was originally set up for maize and soy bean leaves. In experiments preceding this study (not published), chlorophyll contents of beech and oak leaves were measured with Spad and with acetone extraction in a laboratory, and we found that the equation represents the Spad-chlorophyll relationship for these species as well. Prospect-5b inversions of FieldSpec/PlantProbe leaf reflectance measurements confirmed that the chlorophyll contents derived from Spad measurements are in the correct range of values. In Prospect, chlorophyll content is given as μg·cm⁻². By assuming a relative molecular mass for chlorophyll a + b of 900 g·mol⁻¹, the values can be converted from μmol·m⁻² to μg·cm⁻² by applying a multiplication factor of 0.09.

The images recorded are generally of a high signal-to-noise ratio. The spectra are only marginally affected by noise, except for the spectral region beyond 950 nm and in the O₂-absorption peak at 760 nm. The black mats, which have a very low reflectance throughout the spectrum, considerably reduced stray light from the background. With the applied settings, one scan takes about two minutes. Some geometric distortions caused by wind-driven leaf movement during the scanning process are clearly visible. We did not apply any geometric correction to the images.

In each image, the areas covered by the beeches of each pot have been manually identified and marked as regions of interest. Mean spectra of these regions were extracted from the images. Figure 3 depicts the temporal change of reflectance of the dry group pots. At the beginning of the experiment the pots were randomly ordered and assigned to control and dry group. Groups of three pots (except for pots 16 and 20) were imaged together at a time; the plots are grouped according to the image recording configurations. Figure 4 shows the according spectra of the control group pots. The spectra show that many of the dry group trees were heavily affected by a browning of leaves in the last two points in time, with first hints of stress already at the second measurement. Most control group spectra do not change much over time. In the control group, only the spectra of pot 13 show clear signs of yellowing, the remaining spectra show the decline of chlorophyll content only subtly in the wavelength region around 600 nm. Changes in general brightness, like those observed in pots 19 and 14, are likely due to the process of extraction the spectra from the image data with different amounts of shadow and differences in leaf orientation in the regions of interest.

Figure 6(a) shows soil water content measured by TDR probes partitioned for the control group and dry group. The lack of available water for the dry group shows very soon after cutting off water supply. Figure 6(b) and Table 1 show EWT for the two groups. A distinct lag can be witnessed between the quantitative effects of dryness stress on soil and leaf water. EWT stays almost constant during the first four measurements. In the fifth and sixth measurement, EWT of the control group stays constant, while EWT of the dry group decreases. Differences between dry and control group are significant at the fourth measurement and highly significant for the last two measurements. The variance of EWT values is highest in the fifth measurement of the dry group, so is it obvious that not all of the trees suffer from water stress in the same way.

The value range of measured EWT is rather low from the beginning compared to values given in [25] and [39]. In order to confirm these values, EWT was derived from the FieldSpec/PlantProbe reflectance spectra by inverting the Prospect-5b leaf reflectance model. Resulting values correspond well with measured EWT values (Figure 7(a)).

The weather conditions during the experiment are depicted in Figure 8. The trees were kept under a roof, but temperature and humidity were not controlled, so evapotranspiration and the drying rate were affected by the weather. Especially near the end of the experiment, weather conditions were warm and sunny, as reflected in the distinct dryness stress symptoms of the trees.

The PLS regression was able to give an unbiased estimation of EWT with a coefficient of determination of R² = 0.88 (Figure 9(a)). Figure 9(b) shows the PLS regression coefficients of the leaf water content estimation multiplied by the mean reflectance of the training spectra. Leaf water content of a spectrum or a pixel is calculated as the sum of each reflectance value multiplied by its according regression coefficient (inner product) plus an offset. The plot shows the importance of the different wavelength regions for EWT estimation. We see that green wavelengths (around 550 nm) and the red edge region (700 nm) have the largest influence on water content estimation [40]. The NIR plateau region is less important, while the water absorption bands around 970 nm have a high significance [16]. The single negative spike at 757 nm corresponds to the atmospheric oxygen absorption feature. Corresponding spikes can be seen in the reflectance spectra of Figures 3 and 4. We repeated the PLS regression without this wavelength, but no significant difference in the results was detected. The PLS estimation of EWT is quite strongly influenced by the three driest measurements (Figure 9(a)). An alternative PLS model that was trained without these three values reached a coefficient of determination of R² = 0.48, with very similar PLS regression coefficients. We concluded from this that the model used is stable and not solely dependent on the extreme values and, thus, applied the original model to the image data.

Figure 10 shows maps of EWT as estimated by PLS regression for three pots (pots 10, 5 and 3) over time. The general trend towards dryer leaves is clearly noticeable. Some leaves are dry, even in the first images, and some leaves still have high EWT values in the last images. In Figure 2, one can see that not all the trees are equally affected by dryness stress, so these results seem plausible. Areas with negative estimated EWT values and masked background pixels are displayed in black. Some leaves are displayed enlarged to show within-leaf variation. EWT is highest along the leaf veins and lowest in the areas between them. The enlarged leaf in the second subfigure illustrates this. The within-leaf variations of EWT cannot be studied in detail with this kind of imagery, because many leaves are shaded or obscured by others and leaf orientation varies. The trees in the pot shown on the left side of the figures (pot 10) are most affected by dryness stress, the trees in the pot at the right side (pot 3) are least affected. This corresponds well with the spectra in Figure 3. This variation in sensitivity might be explained by the number of leaves and, as a consequence thereof, increased evapotranspiration rates.

Figure 11 illustrates the relationship between buds counted in spring and EWT at the end of the experiment. This supports the assumption that trees with more leaves were more seriously affected by stress symptoms than trees with fewer leaves.

Chlorophyll content declined steadily over time in both the dry group and the control group, as can be seen in Figure 12(a) and in Table 2. Following the drop in EWT, the chlorophyll content of the dry group shows an accelerated decline towards the end of the experiment. While significant differences (p < 0.001) in leaf water content between dry group and control group arise between the 4th and the 5th measurement (between 13 September and 20 September), the chlorophyll contents diverge later, between the 5th and the 6th measurement (between 20 September and 27 September, p < 0.001). The last Spad measurements were only applied on the dry group, after the trees were given water again after the dryness stress experiment. There was a good agreement between the simple ratio calculated from image spectra and the Spad chlorophyll measurements (Figure 12(b)).

There is considerable scatter in Simple Ratio and Spad values of the pots (Figure 12(b)), but, still, the resulting chlorophyll maps are free of image noise and have a reasonable value range and spatial distribution of values. The value range of 100 to 500 μmol·m⁻² total chlorophyll corresponds to 9 to 45 μg·cm⁻².

Figure 13 shows maps of chlorophyll content for one time series of images created from Simple Ratio calculations as described in Section 2.5. Some leaves in Figure 13 are shown enlarged to illustrate the within-leaf variations of chlorophyll content and the sensor’s ability to map them. Because of the vertical extent of the trees, not all of the leaves are in focus, so that the intra-leaf variability is not apparent for all leaves. Some of the leaves seem to have a margin of low chlorophyll content, but we attribute this to edge effects [41]. Chlorophyll variations along and between the leaf veins can be seen in the zoom images. Detailed analysis of within-leaf variation is not feasible with this kind of imagery, because shading and variations in leaf orientation impede comparability; laboratory imaging spectroscopy of single flattened leaves would be a way to study small-scale variations in chlorophyll content.

Chlorophyll content and EWT are correlated in our measurements. Figure 14 shows the correlation of all Chlorophyll content and EWT measurements derived from Prospect-5b inversion of FieldSpec reflectance measurements. The correlation coefficient of 0.72 is highly significant (p < 0.001). Since both chlorophyll and mean EWT decline over time, this correlation might be coincidental. Table 3 shows correlation coefficients and significance levels for each measurement time and for all pots, dry group and control group separately, and demonstrates mostly significant correlations during the whole experiment. The correlation analysis does not reveal whether there is a causal relation between the two variables. Since chlorophyll contents of dry group and control group do not differ significantly until after EWT differences have become significant, we assume that the water shortage causes early senescence and thus chlorophyll decomposition in the dry group.