**1. Introduction**

Today's demands of agricultural cropping systems are high. Agroecosystems have to be highly productive, while the undesirable impact on the environment has to be as low as possible. Resource-conserving methods with a minimum of chemical input are in favor. One vision able to approximate this goal is the use site-specific cropping measures. Site-specific management has the potential to lead to a higher or constant productivity with a constant or reduced input of resources [1]. One group of for site-specific applications are plant protection measures [2].

The spatial occurrence of plant diseases in the field, especially in the early season, is often heterogeneous, while in most cases, plant protection compounds are applied homogeneously onto the crop. This spatial heterogeneity of disease occurrence might lead to diverse fungicide demands that are often not considered. Many diseases first occur in patches before they start spreading in the field. One approach for a site-specific application of plant protection measures might be the application of fungicides to patches of disease occurrence [3–5]. This could prevent or stop disease spreading without applying a fungicide to the whole field [1].

Spectral sensors might be tools able to contribute to site-specific disease management [6,7]. Spectral sensors measure the light reflected from the crop canopy [1]. During pathogen attack and disease development on the crop leaf, diseases establish a spectral fingerprint in the reflected leaf signature [8–10]. These shifts of the signature can be detected using spectral sensors, particularly in the electromagnetic spectrum from 400–2500 nm [11]. Spectral sensors can be divided into hyperspectral and multispectral sensors, depending on their spectral resolution. The number and width of measured wavebands mainly characterize the spectral resolution [11].

Non-imaging hyperspectral sensors average the spectral information over a certain area, while imaging sensors contain the spectral information for each pixel [7]. Hyperspectral imaging sensors (HSIs) provide spectral information in a spatial resolution. Multispectral sensors typically cover the RGB range with an additional NIR band. These sensors are less cost-intensive and the generated data are less complex, but do not cover the broad spectral range like a hyperspectral sensor.

Spectral sensors have been applied on different scales [12]. For field approaches, a hyperspectral camera can be mounted to a ground based vehicle or to a UAV [1,3,11,13]. Depending on the interrogation and measuring setup, each scale can have advantages and disadvantages. On the ground scale, it is possible to detect small features of a few mm through high resolution on close range, while the throughput on the UAV scale is much higher, with still higher resolution compared to satellite imagery [5,14–16]. For field applications of spectral sensors, depending on the scale, the resolution or the measurement time can become a limiting factor. Most field applications for disease detection focused on the calculation of vegetation indices (VIs) [17–19] using multispectral sensors. VIs are developed by accounting certain band ratios to highlight one factor and reduce the impact of another factor [20]. Depending on the wavelength, these indices can be indicators for crop vitality, general crop stress, pigment content, or a specific plant disease [18,21]. Few works have demonstrated an approach for disease detection using imaging hyperspectral sensors under field conditions [10]. This might be because spectral measurements under field conditions are challenging and the complexity of hyperspectral data is higher than multispectral data [1]. The features of multispectral sensors might result in lower image acquisition durations and lower susceptibility to environmental factors during measurements. The image quality of field data in general is influenced by various factors. Beside suitable weather conditions, the field crop species and the disease symptom type are of high relevance for successful measurements. The leaf architecture and disease occurrence on the plant mainly determine the detectability of the disease. Disease presence on lower plant and leaf levels results in a decreased reflected signal. Disturbing weather conditions such as wind and rain can easily obscure spectral images obtained in the field. One elusive and eminently important factor is the illumination. Changing illumination conditions over time, caused by clouds or solar altitude, can lead to uninterpretable data, because spectra of different images cannot be compared with one another anymore [1,3,22]. The detection of diseases on different leaf levels is also challenging because of inhomogeneous illumination conditions through the leaf altitude in the crop stand and upper leaves that cast shadow. These leaves might also be in different developmental stages, and a senescent leaf has to be differentiated from a healthy green or a diseased leaf. The leaf angle to the camera influences the spectral signal. Not least, the image quality is essentially determined by the spatial resolution of the sensor system used. Small symptoms of a disease can only be visualized when the spatial resolution in combination with the measurement distance is appropriate for the desired data quality.

So far, various field measurements on the ground canopy scale of cereal crops have focused on the detection of biotrophic diseases such as yellow rust [3,10,13,23–25], brown rust [18,26], and powdery mildew of wheat [19]. This might be due to the fact that biotrophic diseases are more likely to appear on the upper leaf layers because of wind distribution and a preference for fresh and healthy leaf tissue [27]. Necrotrophic diseases are most severe on lower leaf levels, and therefore more difficult to detect with remote sensors. A detection and quantification of septoria tritici blotch with a hyperspectral radiometer has been demonstrated in the field [28].

The analysis and interpretation of sensor data is crucial for future implementation. Algorithms from machine and deep learning, in combination with suitable sensors and measurement platforms are promising techniques. These methods are particularly able to cope with the number of wavebands provided in hyperspectral data, and can be used for the detection of plant diseases [7,29–32].

This work presents a new approach for field trial studies using innovative and machine learning for a pixel wise detection of crop diseases. A winter wheat trial was conducted in the vegetation period of 2018. The crop was infected with *Puccinia striiformis,* the causal agent of yellow rust (YR). Weekly hyperspectral measurements were performed on the ground-canopy and the UAV scale to monitor the spectral dynamic of crop stands during the vegetation period. Measurements were performed using a mobile field platform with a distance of 50 cm to the crop canopy and with a UAV drone at 20 m height over the plots to work on and compare different scales. Hyperspectral images were captured using a line scanner attached to a linear stage in a measurement booth and a frame-based hyperspectral camera for UAV applications. Field data were preprocessed and normalized, and then analyzed using the supervised classification methods spectral angle mapper (SAM) and support vector machine (SVM) to detect yellow rust of wheat. Additionally, a feature selection was performed on the hyperspectral data to verify the potential for a waveband reduction from hyperspectral to multispectral data for disease detection.
