2.2.1. Field Platform Phytobike

The measurement platform, based on a square steel construction with four wheels and provided by Forschungszentrum Jülich (Jülich, Germany), covered a 3 m wide experimental plot (Figure 3). Sensors for reflectance characteristics, localization systems, power supply, and control of the sensors via a control laptop were mounted to the steel frame. All sensors were variable in height by a moveable aluminum profile construction. In this way, the sensor platform could be adapted to the growth stages of the plants and a constant distance between sensors and crop canopy was enabled. With a weight of around 150 kg, the construction approached the limit of the platforms without steering, and could still be moved by the physical strength of two people.

**Figure 3.** Construction plan of the Phytobike (**top left**) and the final appearance in the field including the cotton diffuser (**top right**). The UAV system used, consisting of a DJI Matrice 600 and a Rikola hyperspectral camera (**bottom left**). Normalization was performed using a 50% grey reference panel (**bottom right**).

As a hyperspectral sensor, the Specim V10E line camera (Specim Oy, Oulu, Finland) was used. The motion required for the Specim V10E camera was realized by a linear stage (Velmex, Bloomfield, USA). Measurements were triggered via the control computer, allowing a flexible reaction to changing light situations by an adapted integration time. The Specim V10E camera measured the electromagnetic spectrum in a range from 400 to 1000 nm with a spectral resolution of 2.73 nm. Sunlight was used as a natural light source. A canvas measuring cabin was constructed to avoid shadows cast by the sensors and equipment of the Phytobike.

## 2.2.2. UAV Measurements

The UAV allowed overview images of whole experiments or at least of parts of the experiment to be collected. Recent technologies have enabled hyperspectral imaging at UAV scale. We combined a UAV DJI Matrice 600 (Da-Jiang Innovations Science and Technology Co., Shenzen, China) with a Rikola hyperspectral camera (Senop Oy, Oulu, Finland) (Figure 3). The Rikola camera measured the reflected light in a range from 500 to 900 nm. The measured wavebands were selectable and spectral resolution was set to 7 nm using 55 wavebands. With flight times of around 20 min, the whole experiment was captured within one battery capacity. For plot observations, a 20 m flight height was selected and the UAV hovered over each plot center for a duration of 10 s. Sunlight was used as a natural light source.

#### *2.3. Data Preprocessing*

This study focused on the information about relevant wavebands as the central outcome. We used a data flow to assess the ability to transfer this information between observation scales (Figure 4). We built two data sets for yellow rust prediction—a classification data set on field scale and a regression data set on UAV scale. Multiple prediction models and feature selection results were derived. In the final step, models were optimized using the selected features, and feature selection information was also exchanged. The resulting four classification models with selected features, two on the field scale (features selected on the field scale and on the UAV scale) and two on the UAV scale (features selected on the field scale and on the UAV scale), were evaluated. This allowed the values of feature selection and, more specifically, the values of feature information obtained at a different observational scales to be evaluated.

**Figure 4.** Underlying workflow of data assessment to compare different prediction algorithms and evaluate the value of information about selected features at different scales. SVM = support vector machine, SAM = spectral angle mapper.

## 2.3.1. Spectral Preprocessing

The derivation of the physical surface property reflectance from observed intensity values is an essential part of hyperspectral image processing. The normalization procedure has to be adapted to the measurement platform and is based on a spectral reference panel with a known homogeneous reflectance in the observed wavelengths. At all scales, the following equation was applied to calculate the reflectance *R* from the observation *Im,* reference *Imref,* and the corresponding dark currents *DCIm* and *DCref*. In the field, the additional *DCref* was omitted for practical reasons.

$$R = \frac{\mathfrak{D} - D\mathbb{C}\_{\mathfrak{D}}}{\mathfrak{D}\_{ref} - D\mathbb{C}\_{ref}},\tag{1}$$

On the ground canopy scale, a 50% spectral reference panel was measured within each image of the line scanner. A separate dark current was observed for the Specim V10E camera before every image. For practical reasons, UAV flight sequences were started with the acquisition of a single dark current, and one image of the reference panel immediately before and after flying the frame-based Rikola camera over the wheat plots. Image quality always suffers from motion of the object to be measured or motion of the sensor. To avoid this, images were taken in conditions as calm as possible on the ground canopy scale. The Rikola camera was hovered for at least 10 images over the reference panel to ensure that image quality was sufficient for data normalization. The use of cross-sensor normalization, e.g., by using a separate spectrometer that continuously logs the incoming light intensity, was tested but was not successful due to a deviating response characteristic between the different sensors.

To remove high frequency noise at the spectral border regions of the Specim V10e, the bands 1–20 (400–450 nm) and 181–211 (910–1000 nm) were excluded from further analysis, resulting in 161 used spectral bands. In addition, a Savitzky–Golay filter using 15 centered points and a polynomial of degree 3 smoothed the data of the Specim V10e.
