*4.2. Hyperspectral Setup Optimisation*

The results from Figures 2 and 3 and Figures 4 and 5 confirmed that the method used for optimising the hyperspectral sensor setup for a wheat canopy can likewise be used to optimise the setup for a potato and leek canopy. The most optimal setups in potato and leek consisted of an ET of 50 ms (Figure 3). There was also a clear trend in the potato SNR values indicating a positive correlation with ET, which is in agreemen<sup>t</sup> with results found in a wheat canopy [19]. For leek, values appeared more variable. The effect of each variable is shown in the PCA factor map plots of Figures 4 and 5. Both figures indicated that height seemed less important for SNR. The angle also showed no clear correlation to SNR in either of the figures. However, the ET vector was clearly congruen<sup>t</sup> with the SNR vector in the potato PCA factor map plots (Figure 4). Together with the values presented in Figure 3, it is reasonable to conclude that ET was the most important factor affecting SNR for this experiment. This is in agreemen<sup>t</sup> with the findings in wheat, where ET was also found to be the most important factor determining SNR values [19]. However, the results in leek were less clear-cut compared to those of potatoes and wheat. This crucial difference means that researchers must always determine the

optimal setup for the crop under observation, before measurements. In our results, the optimal setups consisted of an angle of 8◦, a height of 70 cm and an ET of 50 ms for potatoes and an angle of 17◦, a height of 70 cm and an ET of 50 ms for leek (Figure 3). The optimal setup for leek also registered as the 5th best setup for potatoes. Theoretically, the best measurement setup (of the ones tested in this experiment) for both crops therefore lies at an ET of 50 ms, a height of 70 cm and an o ff-zenith angle of 8◦ to 17◦. However, some practical constraints have to be considered. First, we observed the occurrence of saturation at ETs of 30 and 50 ms in potato and leek, respectively, similar to the results in wheat [19], but there for a much higher ET of 1000 ms. Only the 1 and 10 ms treatments did not show saturation. It is further important to note that a lower ET results in a higher possible framerate. For some measurement conditions, there is a minimum speed at which the measurement needs to be done (e.g., due to time restrictions, operating speed of a treadmill or driving speed of a tractor). This speed correlates to a minimum framerate necessary to obtain a scan of the full sample, which in turn is associated to a maximum possible ET. From Figure 3, it can be concluded that theoretically the optimal height of scanning of both studied crops seemed to be 70 cm. However, since the variability in potato SNR values and to a lesser extent in leek is mainly caused by ET, the choice of measurement height needs to depend on other factors than the SNR values. We propose that measurement height is determined mainly by pixel resolution and scanning width needed in the experimental context. For example, if the main goal is to scan as large a crop area as possible in a limited amount of time (e.g., farmer field scans), then a height above the crop of 110 m is beneficial. However, if the aim of the experiment is to detect small symptoms (e.g., rust pustules) on the leaves of an experimental leek plot in early stages of infection, a height of 30 cm above the crop is preferable, since this will yield a higher image resolution.

The o ff-zenith angle shows no clear correlation to SNR like that of the ET, except for leek in the PC1–PC2 plot (Figure 4). The complex interaction between viewing angle and reflectance has been studied for forest canopies [39]. It was found that for white backgrounds, the reflectance decreased with increasing o ff-zenith measurement angle, while for dark backgrounds the opposite occurred, with increasing reflectance at higher o ff-zenith viewing angles. In a mixed system such as a crop canopy that contains brighter areas (e.g., phytophthora or other wilting symptoms) and also darker areas (e.g., dark spots or shaded areas), it is di fficult to theorise on the e ffect of viewing angle on reflectance and SNR. Results in pine forests showed that there was a specific angular e ffect on the reflectance of the red and red edge bands for coniferous trees, possibly due to their canopy structure [40]. Such an angular e ffect due to canopy structure could contribute to the di fference between the results for potato and leek, since angle only seemed to contribute to SNR for leek canopy measurements. This is supported by the clear di fference between the oblique growth pattern of a leek canopy, where angular di fferences are significant between leaves, compared to the relatively homogenous growth pattern of a dense potato canopy, where angular di fferences might be more easily averaged out. It has further been shown that the angle significantly a ffects certain canopy measurements, for example of vegetation indices [41]. This is relevant when performing image analysis on, for example, single leaves. In such case, the e ffect of geometry on reflectance can be modelled to improve data analysis [41–44]. The same applies to satellite imagery, where the modelling of the solar o ff-zenith angle is crucial [45]. However, for practical applications in proximal crop sensing, such complex crop geography modelling techniques are often omitted, and it is assumed that the e ffect can be neglected, or vegetation indices are used that are resistant to these e ffects [41,46]. We therefore propose that the measurement angle needs to be estimated based on the disease of interest. Ideally, scans should be taken perpendicularly to the plane, in which symptoms occur, to maximize the chance of scanning the infected area. If the disease is for example known to manifest on the bottom of the stem, a 0◦ off-zenith angle will have no chance of detecting it, as opposed to a 17◦ angle, which makes it possible to scan the lower canopy. Results for powdery mildew detection in grapes also supported the use of higher o ff-zenith angles [35]. This leads us to advise an o ff-zenith angle of 17◦ to detect symptoms on the bottom of the leek stem and lower leaves. Phytophthora damage tends to appear at leaf tips, but also on the base of the leaf where spores are splashed onto the plant from the soil [47]. To detect symptoms on lower leaves of the potato canopy, an angle of 17◦ is again advised, especially in dry growing conditions that cause the leaves to sag, causing the symptoms to be in the vertical plane perpendicular to the soil.

Although the use of artificial lights has certain advantages (see subsection 3 of the discussion), it also increases the risk of saturation, especially at low measurement heights. The choice of the optimal measurement setup is therefore not only based on the best SNR but is far more complex and should be discussed from practical perspective, taking into account the specific case of each disease or crop under observation. A balance needs to be found between high ET, resulting in better SNR values, while taking into account saturation, pixel size, measurement speed and the structure of the pathosystem. Specifically, for leek, white tip disease (Phytophthora porri) symptoms can easily cause saturation at low scanning heights because of their proximity to artificial lights, even at low ETs. The trade-o ff between the risk of saturation (at high ETs) and the risk of noisy data (at low ETs) can partly be overcome by spectra preprocessing techniques, which can, to some extent, deal with noisy spectra [48]. It is important to note that the reflectance values in the visible part of the spectrum are much lower compared to the near-infrared part of the plant canopy spectrum [49]. This means that saturation could primarily occur in the near-infrared part of the spectrum, while a low ET and subsequent noisy data could occur more easily in the visible part of the spectrum. Depending on which part of the spectrum is mainly of interest, the saturation/noise trade-o ff might be di fferent. For proximal disease detection in experimental conditions for leek and potato, we advise a height of 30 cm (instead of the SNR optimum of 70 cm) above the crop canopy, to maintain a high resolution. Because of the risk of saturation at this height, an ET of 1 ms is advised for both leek and potato, instead of the SNR optimum of 50 ms (for both leek and for potato). This needs to be increased if spectra appear "flat", with minor or without features. However, saturation should be absolutely avoided. Since no clear influence was observed for the angle, it is recommended to use the 17◦ angle because it can measure symptoms on lower canopy parts. This was the theoretical optimal angle for a leek canopy, but for a potato canopy the theoretical optimum was an 8◦ angle.

### *4.3. E*ff*ect of Artificial Lighting on Hyperspectral Measurements of a Crop Canopy in the Field*

To explain the phenomenon that the added light seemed to decrease reflectance of spectra collected from a canopy on a cloudy day (Figure 6B), the spectral profile of the white reference was investigated (Figure 7). At wavelengths up to +/-470 nm, the di fference between the light on/off white reference reflectance curves under cloudy conditions (Figure 7B) was minimal, similar to sunny conditions (Figure 7A). At higher wavelengths, the di fference increased. Without additional light, the white reference reflectance curve under cloudy conditions was skewed compared to the curve under sunny conditions (due to the e ffects of cloud cover), leading to an inappropriate correction at high wavelengths. The corrected spectrum of wavelengths > 470 nm was 'stretched' compared to those at lower wavelengths. This is due to lower white reference reflectance values, which increases the final corrected value according to Equation (1). This explains why in Figure 6B, the added light under cloudy conditions caused a decrease in reflectance compared to the light-o ff scenario. Turning on the artificial light increased the white reference to such a degree that it counteracted the increased reflection from the crop canopy (due to added artificial radiation), resulting in ultimately lower reflectance values. If there would have been no artificial illumination during experiments, the skewed white reference values on cloudy days would cause the final reflectance in the >470 nm range to be relatively stretched (as shown in Figure 6B) compared to the same target measured on a sunny day (Figure 6A), independent of the crop health status. The added light helped to counteract this problem.

After normalisation, the shapes of the light on/off canopy reflectance curves are similar under sunny conditions (Figure 6C), with only a small di fference in the NIR part of the spectrum that could be amended by further preprocessing [50]. The di fference between light on/off canopy reflectance curves is much more severe under cloudy conditions, especially around 680 nm. This is an important region for disease detection, so it is essential that any change in reflectance is the result of disease, rather than cloud cover variation [11]. The apparent increase in Figure 6D between 600 to 680 nm could be misinterpreted as an increased 'red-orange' colour, which could be interpreted as rust disease symptoms (results not shown). The addition of artificial light helps counteract this e ffect to some extent, but it is still important to keep this in mind for further data analysis. This area, especially the red colour band at 680 nm, is well documented in literature as being a spectral feature indicating chlorophyll absorption, making it one of the most important features in crop health sensing [19]. The drop in measured reflectance at +/-930 nm appears in both the lights on and lights o ff curves during cloudy conditions, which suggests that this is a feature, rather than purely the result of variations in solar radiation intensity due to cloud cover (Figure 6D). The fact that this drop occurred only in cloudy conditions lead us to look at the absorbance spectra of cloud cover reported in literature [51]. These authors reported a significant absorption band around 940 nm for stratus clouds, which could possibly explain the decrease in reflectance. This could indicate that the artificial light was not strong enough to compensate for the absorption caused by cloud cover in this spectral region.

Another important aspect of the artificial light is that the angle of reflectance is di fferent compared to that of natural solar radiation [18]. During sunny conditions, the light strikes the crop at a certain angle, depending on the solar radiation. With the artificial light, the light strikes the crop from both sides (because two lamps are used in the present work), at a constant angle to the sensor. This means that the light not only counteracts the e ffect of cloud cover but also provides a constant source of illumination that helps mitigate the e ffects of changing solar angle. It is therefore advisable to use as much artificial lights as possible, to reduce the e ffect of solar radiation variation.

### *4.4. E*ff*ect of Artificial Lighting on Thermal Measurements of a Crop Canopy in the Field*

If the crop temperature would increase significantly during sensor movement due to the added radiation, it would be most apparent on the top leaves, which receive most of the radiation. During sunny conditions, hot spot formation was only apparent at the edge of the crop row, where the sensor box stayed stationary for more than 10 seconds (Figure 8.1B). This indicates that even though there was a temperature di fference during these measurements, it was mainly caused by absorption of radiation by the soil, not the crop. Otherwise, hot spot formation would occur on the top leaves during measurements (and not only when the box was stationary). Note that this implies that correcting for the temperature di fference caused by artificial light is di fficult because not every part of the crop row heats up equally. The assumption that bare soil is mainly responsible for the temperature di fference between light on/off is also supported by the fact that average measured temperatures varied between the edge and the middle of the crop row, where the soil-to-crop ratio is di fferent (Figure 9). An important observation is that during cloudy conditions (Figure 8.2), there was no apparent temperature increase compared to 1 ◦C for sunny conditions (Figure 8.1). There was even a small apparent decrease (0.5 ◦C or less) after turning on the lights for some images during cloudy conditions. The temperature decrease is possibly due to the e ffect of variations in cloud cover, which are di fficult to record during the time of one scan. The lack of temperature di fference could also be explained by the dense weed cover that covered the darker soil later in the growing season during the measurements under cloudy conditions. This suggests that it is possible to use halogen lamps in combination with a thermal sensor for weed or crop cover assessment.

During cloudy conditions, hot spot formation in the middle of the crop row was observed during measurements, not only when the sensor box was stationary (Figure 8.2A). The leek crop was older during this measurement, showing signs of wilting on older leaves in places where the leaves were bent or cracked. It was in here that hot spot formation occurred, even when the sensor box passed over the crop in a matter of seconds. This indicates that artificial light interacts di fferently with diseased or damaged and healthy parts of the crop and that this di fference can be observed with thermal cameras. This feature could assist in disease detection using thermal cameras in addition to the fact that these sensors can measure the temperature di fferences due to evapotranspiration. It is also important to note that the interaction between artificial light and damaged crop areas occurred even over a matter of seconds, which is much faster than the rate at which evapotranspiration changes occur [52]. This could provide new research opportunities, for example by placing thermal cameras with artificial lights at the back of weeders for detecting mechanical damage after passage, which would take longer to show if no artificial light is present.
