Practical recommendations and limitations for pushbroom hyperspectral imaging of tree stems

.


Introduction
Optical airborne and satellite remote sensing platforms can provide cost-efficient means to gather spectral information about forests from local to global scale.This information is particularly valuable for, e.g., forest ecology, management, and health assessment, if the reflected signal is interpreted correctly (Nilson et al., 2003).The total reflectance of a forest is formed by a mixture of forest elements that contribute in varying fractions at different wavelengths of the spectrum.In addition to the foliage and the forest floor, the woody elements (e.g., stems and branches) of the forest canopy affect the total reflectance considerably.Gower et al. (1999) reported that the woody elements of a forest can contribute 5% to 35% of the total plant area.Studies utilizing forest reflectance models have shown that the contribution of woody elements to the total forest reflectance is significant especially in the near-infrared (NIR) wavelength region (Halme and Mõttus, 2023;Hovi et al., 2022;Widlowski et al., 2014).Despite previous knowledge, the contribution of woody elements like stem bark is often overlooked in physically-based remote sensing models, ultimately leading to biases in our interpretations of remote sensing data.
The reason for ignoring the spectral contribution of stem bark in physically-based remote sensing is the sparse availability of reference data.The number of studies that have collected spectra of stem bark directly in the forest are limited because the measurements are tedious, costly, and there are no standardized methods or protocols on how to measure vertical tree elements.In situ spectral measurements of stem bark have been conducted with conventional point spectroradiometers (e.g., Girma et al., 2013;Hadlich et al., 2018;Lang et al., 2002), but recent technological advancements are bringing novel, portable, and close-range capable sensors, such as small pushbroom imaging spectrometers (e.g., utilized by Juola et al., 2022a), for consumer and research use.The benefit of novel small-sized imaging spectrometers is that they allow users to simultaneously collect spectral and textural information of surface materials in the field.Highly detailed spectral and textural information of stem bark could be utilized with, e.g., 3D terrestrial laser scanning data to produce realistic forest stand reconstructions needed to facilitate the calibration and validation of remote sensing data (Calders et al., 2018).As new sensors are introduced, it highlights the importance of investigating the strengths and limitations of different measurement techniques.Guidelines for field measurements with point spectroradiometers exist (as reviewed by Milton (1987) and Milton et al. (2009)), but quantitative knowledge on the factors influencing the measurements under forest canopies and with modern pushbroom imaging sensors is lacking.
Acquiring close-range spectral images of forest elements in their natural state holds many practical challenges.One of the largest challenges is that the measurements often must be conducted under varying illumination conditions.In this short communication, we report how measurement conditions affect the in situ collection of stem bark spectra.The data for this study were collected in forest environments with a cutting-edge pushbroom hyperspectral camera designed for portable measurements.The two research questions to be examined are: 1) How do varying measurement conditions influence the quality of stem bark reflectance data?, and 2) What practical recommendations can be made to overcome the influence of varying measurement conditions?

Study sites
We acquired hyperspectral data cubes of Norway spruce (Picea abies  We selected three trees of different sizes per species from mainly monospecific stands (except for a few trees within the stands) representing varying stand density, canopy closure and plant area index (Table 1).For each tree we collected 19 hyperspectral data cubes per day from a fixed sideways-looking view angle (perpendicularly to the northern side of the stem so that the camera viewed from north to south).The hyperspectral data cubes were acquired every 30 minutes, starting from 08:30 to 20:00 (UTC + 03:00).On each measurement day, data were not acquired between 11:30-12:00 and 17:30-18:30 due to meal breaks.
To provide background information on forest canopy structure, we used digital hemispherical photographs (DHP), taken in diffuse illumination conditions to estimate the effective plant area index (ePAI) and canopy closure (CC) for each forest stand (Table 1).The DHPs were taken with a Nikon D5000 DSLR camera with a geometrically calibrated Sigma 4.5 mm 1:2.8 DC HSM Circular Fisheye lens.ePAI estimates the ratio of plant area to ground area and CC estimates the proportion of the sky hemisphere obscured by the canopy.We took five DHPs per stand, four from the cardinal compass directions at 5 m distance, and one next to the sampled tree.We used the algorithm by Nobis and Hunziker (2005) and the blue color channel to automatically threshold DHPs into binary images.Finally, we used the averaged gap fractions and the equations from the LAI-2200 instrument manual (LI-COR, 2012) to calculate the stand-level ePAI and CC.

Hyperspectral imaging
The hyperspectral data cubes were acquired with a pushbroom imaging spectrometer (Specim IQ, Specim Ltd., ser. nr. 190-1100152) which was attached to a mounting setup designed for vertical tree stems (see Juola et al., 2022a) (Fig. 1).The mounting setup from Juola et al. (2022a) was modified slightly to allow the white reference to be simultaneously present in the camera's field-of-view during data collection (Fig. 1).The direction of the pushbroon scanning in the setup was from left to right.The trees were imaged from approximately 1.3 m height and at a 25 cm distance.The spectral range of the Specim IQ is 400 nm -1000 nm and the spectral resolution (FWHM) is 7 nm.The field of view of the Specim IQ is 31 • × 31 • and the raw data cubes are shaped 512 × 512 × 204 (rows/columns/bands).The raw data cubes were processed into reflectance with a commonly used formula (Juola et al., 2022a) that removes the dark current and normalizes the images to the white reference.The white reference used in this study was a calibrated Spectralon® panel that has a nominal reflectance of 99%.Reflectance factor of the panel was taken into consideration in the formula when processing the raw data cubes into reflectance factors.The white reference covered the entire width of the image and filled a total of approximately 15% of the bottom of the image area (Fig. 1).Each image was normalized to the average white reference signature within the image to obtain hemispherical-directional reflectance factors (HDRF, Schaepman-Strub et al., 2006).Integration times were set manually for each measurement to maximize the dynamic range without saturating any pixels (Table 1).Due to low signal-to-noise ratio between 400 nm and 415 nm and reported instability of the signal between 925 nm and 1000 nm (Behmann et al., 2018), the wavelength range of the data was clipped to include 415 nm -925 nm.Subsequent analyses after processing the raw data cubes into reflectance images were made with the pixels representing the bark area only.

Radiation and meteorological measurements
To characterize variation in environmental conditions during the measurements, we used data on wind speed, global shortwave solar radiation, photosynthetically active radiation (PAR), diffuse PAR, and cloud base height provided by the SMEAR II research station (Aalto et al., 2022).
The horizontal wind speed (at 33.6 m height) was measured with a Thies 2D Ultrasonic anemometer.Global shortwave solar radiation in the wavelength range of 0.3 μm -4.8 μm (at 35 m height) was measured with a Middleton solar EQ08 pyranometer.Photosynthetically active radiation (PAR) in the wavelength range of 400 nm -700 nm (at 35 m height) was measured with a Li-Cor Li-190SZ quantum sensor.Diffuse The temporal frequency of measurements for all variables was 1 minute.We used a five-minute window around each measured image to calculate the average wind speed and shortwave global radiation.In addition, we calculated the cloud occurrence (percentage of clouds detected above the ceilometer within the time window), and a cloud index (diffuse PAR divided by total PAR) for the same time window.These four proxy variables were hypothesized to be related to the radiation conditions (e.g., quantity, temporal variability, and direct and diffuse fractions of solar radiation) during the image acquisition, and could have affected the image quality.

Influence of measurement conditions on the quality of hyperspectral images
In typical non-diffuse illumination conditions (i.e., direct solar beam dominating the top-of-canopy irradiance), the likelihood of obtaining valid field measured hyperspectral images of stem bark with a fixed view pushbroom camera setup was heavily dependent on the solar azimuth and zenith angles.We considered images to be valid when the white reference panel and the bark surface received equal-like illumination as per the definition of relative reflectance factors (Schaepman-Strub et al., 2006).Visual classification revealed that the number of total valid images per acquisition time increased towards mid-day (12:30 to 14:00 UTC + 03:00) (Fig. 2).Close to mid-day, the Sun was either mostly or fully obstructed by the stem itself so that nearly all images taken were valid.In contrast, diffuse illumination conditions during a fully overcast sky promoted stable results even for measurements in difficult solar azimuth and zenith angles, such as morning or evening measurements (Fig. 3, Example 1).
We observed that temporal changes in irradiance occurred at different time frames during image acquisition with the pushbroom sensor (Fig. 3, Example 2).These included short period (less than a second, i.e., from one sensor array integration to another) and longer period fluctuations (seconds to minutes, i.e., across the entire image acquisition time).As expected, considerable changes in irradiance occurred when clouds passed in front of the Sun or when the canopy swayed back and forth.This meant that the data quality of images decreased as the variation of irradiance increased, which could be seen as striping and uneven illumination within the image area (Fig. 3, Example 2).In the non-optimal measurement times and under partly cloudy or clear sky, these fluctuations were often unpredictable as clouds or canopy elements moved in and out of the direct solar beams.However, the most significant changes in irradiance occurred spatially between the white reference panel and the bark surface, which were ultimately used to calculate the reflectance quantities.For example, in certain illumination angles and non-diffuse conditions, strong direct solar radiation and shadows manifested on the rough curved surface of the stem bark but not equally on the flat white reference panel (Fig. 3, Example 3).It was also possible for the white reference panel to receive direct spot-like illumination through the canopy which was not present on the stem bark (Fig. 3, Example 4).
For all tree species, we observed diurnal variability in the shape and overall level of the average spectra (Fig. 4).The variability was linked with the validity of image (Fig. 4).If invalid images (Fig. 4 orange dashed lines) were removed, the per day variations left for the average spectra of valid images were fairly small for all species (in the range 0.6% to 5.9%, Fig. 4 black solid lines and gray shaded regions).This highlights that images obtained from the side of stem that is opposite to the Sun, or when the clouds block the direct solar radiation, produce comparable results even for differing solar zenith and azimuth angles.The effect of specific view-illumination geometry (i.e. that produces both sun-flecks and shadows present in the image) on the spectra was the largest in visible (VIS) and red-edge regions, and the smallest in NIR (Fig. 4, orange dashed lines).A likely explanation is that the strong NIR scattering by tree canopies reduces the irradiance differences between sun-flecks and shadows, and thus the irradiance is spatially more homogenous in NIR than in VIS.There is a possibility that the magnitude of the changes in the average reflectance spectra were also related to the surrounding canopy structure and the different bark textures between the species, though our dataset was not large enough to determine this quantitatively.We observed that, e.g., forest stands that are dense and have closed canopies have a higher likelihood of preventing direct solar radiation from reaching the bottom-of-canopy.
Finally, our objective was to analyze how varying meteorological and radiation conditions affected the quality of the measurements.For each measurement day, we examined how the within-image (i.e.bark area only) coefficient of variation (CV) depended on the proxy variables that described the radiation conditions.The within-image CV is an average over all wavelengths.Uncertainties in the meteorological and Fig. 4. Average reflectance spectra (black solid lines) and the ±2 standard deviations (shaded gray region) for all valid measurements per day and species.Dashed orange lines are the reflectance spectra of invalid images classified as having a specific view-illumination geometry between the white reference panel and the sample surface that produced both sun-flecks and shadows present in the image.radiation variables were presented as their standard deviations within the five-minute observation windows (Figs. 5, 6, and 7).The images with direct solar radiation had outlying CVs, hence negligible correlations between CV and the proxy variables were observed when analyzing all images together.Therefore, the correlation analysis was made with images that were classified as valid.First, we analyzed how changes in shortwave global radiation are related to spectral image quality.Pearson's correlation coefficient revealed primarily negative correlation (r = −0.04 to −0.75) between the CV of bark reflectance factors and the average shortwave global radiation (Fig. 5A, 6A, and  7A).The negative correlation could be explained by the decreasing signal-to-noise ratio (despite the optimized integration times) when the shortwave global radiation decreases, or by potential changes in the directional distribution of incoming radiation associated with the changes of global shortwave radiation.Depending on the surface roughness of the bark sample, the changes in directional distribution of incoming radiation could influence the contrast between the bright (well-illuminated) and the dark (shadowed) areas of the bark sample.
Next, we looked at how cloudiness influences the image quality using a cloud index and cloud occurrence data.Pearson's correlation coefficient between the CV of bark reflectance factors and the cloud index showed primarily weak to strong positive correlation (r = 0.36-0.78)(Fig. 5B,  6B, and 7B).Noteworthy is that for a few of the samples, e.g., the silver birch that was measured on 29.6.2021,there was not enough variability in the cloud index to be able to generalize the result even though a linear regression was fitted (Fig. 5B).The relationship with cloud occurrence and the CV of bark reflectance factors was too complex to describe linearly (as seen from, e.g., the clusters at 0% of cloud occurrence in Fig. 5C  and 7C).Favorable measurement conditions could be achieved when cloud occurrence was either 0% or 100%, but the CV of bark reflectance factors still varied, probably because of the variability in global shortwave radiation and illumination geometry depending on the time of day (Fig. 5C, 6C, and 7C).Also, while clouds can provide more diffuse measurement conditions and block direct solar radiation, constantly varying cloud cover (i.e., cloud occurrence values between zero and one) will introduce variation into the pushbroom image data.Finally, Fig. 5. Relationship between the average coefficients of variation for bark reflectance factors on three separate measurement days (29.6.2021, 7.7.2021, and 13.7.2021)for silver birch and A) the average shortwave global radiation, B) the average cloud index (estimated as the ratio of diffuse photosynthetically active radiation to total photosynthetically active radiation), C) the cloud occurrence (estimated as the percentage of clouds detected above the ceilometer), and D) the average wind speed within a five-minute window around each valid measurement.The horizontal lines are ±1 standard deviations of the meteorological and radiation variables during the five-minute observation windows.In C), the correlations for the valid measurements on 29.6.2021(orange points) were not calculated because the cloud occurrences were 0% (marked as "n/a" in the legend).we looked at wind speed which also had a scattered relationship with the CV of bark reflectance factors (Fig. 5D, 6D, and 7D).While the canopy swaying in the wind was noted to produce variation in some individual images (Fig. 3 Example 2), the wind speed did not produce a meaningful correlation.Wind speed and cloud occurrence most likely have nonlinear relationships with the CV of bark reflectance factors, but the number of observations was not sufficient to determine the shapes of the relationships.Overall, for the valid images, the difference between the minimum and maximum CV of bark reflectance factors was 2%-19%, depending on the measurement day.According to the r values, up to 61% of this variation could be explained by the meteorological and radiation variables (r = 0.78 corresponds to r 2 = 0.61) (Fig. 5).

Practical recommendations for pushbroom hyperspectral imaging of tree stems
Finally, based on experiences gained during the field campaigns of this and previous studies (Juola et al., 2022a), we can make practical recommendations for pushbroom hyperspectral imaging of tree stems.
With good measurement practices it is possible to mitigate many of the random fluctuations and uncertainties that arise when imaging tree stems in difficult field conditions.
(1) Our first recommendation is that the pushbroom imaging platform should be stable and robust.A fixed or fixable geometry between the camera, white reference panel, and the stem allows for the acquisition of repeatable and comparable measurements with minimal distortions (Fig. 1).Alternatively, a darker reference panel could be used in the future to improve the radiometric resolution and to ensure that a greater portion of the dynamic range of the imaging spectrometer is used during measurements.
The measurement geometry also determines the reflectance factors that are acquired (Schaepman-Strub et al., 2006).Hence, the camera and other equipment should be placed the same way (relative to the Sun) for all measurements in the same campaign.
(2) Second, we recommend paying attention to the overall measurement set-up by ensuring that scattering from nearby equipment and clothing is minimized.For example, standing too close during the image acquisition, e.g., in bright or reflective clothes can influence the data quality negatively (Kimes et al., 1983).Measurement set-ups are often painted matte and black to reduce its reflectivity, but other equipment should be placed far enough away to ensure they do not contribute to the measured quantities.
(3) Third, we recommend measuring from the side of the tree stem that is opposite to the Sun to avoid direct solar radiation that can distort the spectra (Fig. 4).More specifically, we recommend measuring so that the stem obstructs the Sun, i.e., measuring close to mid-day if the fixed view is from north-to-south.The same principle can be applied for other solar azimuths.In addition to decreasing the probability of sun-flecks and direct solar radiation, measurements on the side of the stem that is opposite to the Sun can help to reduce the effects of solar zenith angle, canopy structure, and varying weather conditions.Averaging spectral signatures in post-processing can further account for heterogeneity and noise caused by, e.g., the scanning process of the pushbroom imaging spectrometer (Behmann et al., 2018).
(4) Fourth, we recommend performing measurements in stable meteorological and radiation conditions.Taking into consideration our previous recommendations, measurements during cloudless clear-sky days can produce high signal-to-noise ratio data and low integration times (i.e.fast measurements).However, during clear days extra attention should be paid to random fluctuations in illumination caused by, e.g., clouds obstructing the Sun or sun-flecks through the canopy during the measurement (Fig. 3, Example 2).However, diffuse overcast days with clouds can extend the time window for measurements, making it a practical choice for acquiring hyperspectral images of stem bark.5.7.2021, 8.7.2021, and 14.7.2021) for Norway spruce and A) the average shortwave global radiation, B) the average cloud index (estimated as the ratio of diffuse photosynthetically active radiation to total photosynthetically active radiation), C) the cloud occurrence (estimated as the percentage of clouds detected above the ceilometer), and D) the average wind speed within a five-minute window around each valid measurement.The horizontal lines are ±1 standard deviations of the meteorological and radiation variables during the five-minute observation windows.In C), the correlations for the valid measurements on 14.7.2021(blue squares) were not calculated because the cloud occurrences were 0% (marked as "n/a" in the legend).

Links to airborne and satellite remote sensing of forests
Acquiring close-range hyperspectral reflectance images of stem bark in the field can provide detailed insights into the physical and biophysical properties of individual trees and species.Extensive in situ spectral libraries can be used to build a comprehensive understanding of the spectral diversity of forests needed in remote sensing applications and biodiversity mapping around the globe (Asner and Martin, 2016;Rocchini et al., 2022).Despite the limited availability of in situ spectral observations of stem bark, linking such data with airborne or satellite remote sensing has already been shown to improve the performance of, for example, forest reflectance models (Halme and Mõttus, 2023;Hovi et al., 2022;Malenovský et al., 2008).In addition, information on stem bark reflectance has been utilized with 3D terrestrial laser scanning data to construct structurally and radiometrically accurate virtual forests for radiative transfer model simulations (Calders et al., 2018).Woody elements have also been reported to have an impact on the retrieval of leaf area index (LAI) (Myneni et al., 1997) and chlorophyll content (Verrelst et al., 2010).In situ measurements of stem bark spectra can also help to evaluate the contribution of woody elements to tree canopy reflectance (Kuusinen et al., 2021).The spectral signatures and texture of stem bark for temperate and boreal tree species have also been found to be promising for in situ tree species identification (Juola et al., 2022b).These examples show that remote sensing applications are supported by faster, more accurate, and more consistent methods developed for collecting spectral field reference data.

Conclusions
Physically-based remote sensing methods can utilize spectral information of stem bark.The benefit of the bark data can be maximized by improving the accuracy of in situ spectral measurements.In situ measurements are challenged by varying illumination conditions, which in turn can have a significant effect on the quality of the data.We have shown that simply taking into consideration the meteorological and radiation variables with a well-planned measurement set-up, it is possible to improve the precision of in situ collected spectra of stem bark.We made four practical recommendations for pushbroom hyperspectral imaging of tree stems in field conditions.

Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Fig. 1 .
Fig. 1.Example of the measurement setup when attached to a tree stem.The schematic illustrates the setup in more detail.The dimensions shown are in centimeters.

Fig. 2 .
Fig. 2. Histogram of the number of images per measurement time and species that were classified as valid, i.e., in which the white reference panel and the sample surface received equal illumination.

Fig. 3 .
Fig. 3. Four visual examples of commonly observed phenomena when acquiring reflectance images of stem bark in natural illumination conditions: Example 1) diffuse and stable irradiance conditions during a cloudy morning.Example 2) temporally fluctuating irradiance (e.g., due to wind moving the tree crowns) in clear blue-sky conditions with the Sun behind the stem.Example 3) sun-flecks and shadows on the curved stem caused by direct solar radiation from the side.Example 4) white reference panel received a random spot of direct solar radiation through the canopy.The narrow images for each example visualize the percentage change relative to the mean irradiance incident on the white reference panel.The corresponding larger images are RGB composites of the original hyperspectral reflectance images with the white reference panel visible on the bottom of the image.

Fig. 6 .
Fig.6.Relationship between the average coefficients of variation for bark reflectance factors on three separate measurement days(1.7.2021, 6.7.2021, and  9.7.2021)  for Scots pine and A) the average shortwave global radiation, B) the average cloud index (estimated as the ratio of diffuse photosynthetically active radiation to total photosynthetically active radiation), C) the cloud occurrence (estimated as the percentage of clouds detected above the ceilometer), and D) the average wind speed within a five-minute window around each valid measurement.The horizontal lines are ±1 standard deviations of the meteorological and radiation variables during the five-minute observation windows.

Fig. 7 .
Fig. 7. Relationship between the average coefficients of variation for bark reflectance factors on three separate measurement days(5.7.2021, 8.7.2021, and  14.7.2021)  for Norway spruce and A) the average shortwave global radiation, B) the average cloud index (estimated as the ratio of diffuse photosynthetically active radiation to total photosynthetically active radiation), C) the cloud occurrence (estimated as the percentage of clouds detected above the ceilometer), and D) the average wind speed within a five-minute window around each valid measurement.The horizontal lines are ±1 standard deviations of the meteorological and radiation variables during the five-minute observation windows.In C), the correlations for the valid measurements on 14.7.2021(blue squares) were not calculated because the cloud occurrences were 0% (marked as "n/a" in the legend).

Table 1
List of measured trees, species, and structural characteristics of the surrounding forest stands.DBH is the diameter at breast height (1.3 m), CC is the canopy closure as a decimal fraction, and ePAI is the effective plant area index.