Hyperspectral imaging proved to be highly suitable for the detection, identification and quantification of fungal diseases on the leaf level. Each disease influences the spectral reflectance of sugar beet tissue in a specific way resulting in disease-specific spectral signatures. Similar effects have been described previously for foliar and soil-borne diseases of sugar beet by Mahlein et al.  and Hillnhütter et al.  using non-imaging hyperspectrometry.
Since the portion of a signal from diseased tissue in a mixed signal depends on disease severity, the sensitivity and specificity of non-imaging spectroradiometers is limited. Especially at low disease severities, spectra are based on high percentage of reflectance from healthy tissue and only a low portion of symptomatic tissue causing changes in the spectrum. Diseased plants or leaves may be detected by using non-imaging spectrometry, however, only imaging techniques with high spatial resolution - i.e. operating in proximity to the objects - allow for the detection, identification, and quantification of disease symptoms. Disadvantages of non-imaging spectroradiometer, including the separation of mixed infection with two or more diseases on the same leaf or plant, can be overcome by the use of HSI technology.
A pixel-wise attribution of disease-specific symptoms and healthy tissue is conducive to observe spectral reflectance patterns of foliar diseases in detail. Some disease symptoms can only be distinguished from other diseases and stresses when hyperspectral imaging with high spatial resolution is used [17, 14]. The detection limit of a non-imaging spectroradiometer was 10% diseased leaf area for CLS and powdery mildew and 20% for SBR, respectively . In contrast, single symptoms can be detected and identified by using HSI systems, since pure signals from pixels of diseased tissue are recorded.
High spatial resolution is crucial in particular for the detection of leaf diseases with discrete, roundish symptoms like CLS or SBR. Spatial resolution of the hyperspectral camera used in this study provided information even on subareas of disease symptoms. Nevertheless, the tiny uredinia of U. betae and limited spatial resolution of the sensor resulted in a high amount of mixed pixels in SBR experiments. Depending on the shape of the symptoms, pixel size should be smaller than the object of interest by a factor of 2 to 5 [23, 24]. This rule from remote sensing restricts the sensing of plant diseases to proximal sensing technologies.
The sugar beet diseases differed in their temporal and spatial development as well as in their effects on plant tissue associated to reflectance characteristics. The spectral impact of sugar beet diseases on leaf reflectance was previously described in detail by Mahlein et al.  On hyperspectral imaging cubes along transects through CLS symptoms revealed a continuum from healthy tissue, over newly colonized tissue, discoloured (reddish-brown) tissue, chlorotic cells, dying cells and dead cells in the centre of mature associated with characteristic changes of reflectance in the VIS and NIR, which is especially sensitive to modifications of the tissue structure [25, 26]. Boyer et al.  described similar effects in senescent leaves of the northern pine oak.
The biotrophic pathogens U. betae and E. betae are less destructive; both pathogens largely rely on the integrity of host cells and functionality of metabolism of their host plant. Structural changes of infected leaf tissue and modifications in pigment content were smaller than for CLS-diseased leaves and resulted in only slight changes in VIS and NIR reflectance. The number of chloroplasts was not visually affected at the time of the appearance of mature symptoms. The whitish mycelium of E. betae on the surface increased tissue reflectance over the full range of the sensor. An unambiguous detection of powdery mildew in early stages is challenging since the dust-like cover results in a parallel shift of reflectance with minor influence on the shape of the reflectance curve. The reddish-brown urediniospores of Uromyces betae, in contrast, influenced tissue reflectance similar to the reddish-brown margin of CLS symptoms (minor decrease from 450 to 500 nm, increase from 550 to 700 nm, decrease in NIR, Figure 4). High concentrations of carotenoids and melanin-like pigments, causing the characteristic brown-orange colour of urediniospores are well documented for many rust fungi . As stated by Gitelson et al.  carotenoids and chlorophyll have overlapping absorption bands in the blue range around 520 nm. Cercospora leaf spots cause an increase in reflectance from 400 to 550 nm whereas sugar beet rust causes no increase in reflectance in this range. Reflectance is constant with roundabout 0.05%/100; it is assumed that the carotenoids described for sugar beet rust urediniospores counteract the effect of the chlorophyll loss. Nevertheless, the small size of SBR colonies impeded the detection in early stages or at low disease severity.
Spatial patterns of discrete symptoms of sugar beet diseases could be investigated by pixel-wise assignment of spectral signatures. Modifications of spectral reflectance at different developmental stages were displayed in spectral signatures of different subareas of the symptoms. For instance reflectance of new, immature symptoms was similar to that from the margin of fully developed lesions. The results for powdery mildew and SBR generally confirm the principle that maturing, but still growing disease symptoms include all developmental stages so far.
Specific effects of diseases, disease stage, and the impact of disease severity on spectral characteristics of plants are complex, but may allow for new insights into host-pathogen interactions . Similar to Ustin and Gamon , who classified different plant functional types based on morphological and physiological traits into 'optical types' by reflectance measurements, spectra of subareas of infected tissue categorised during disease development in a similar way. Hyperspectral imaging clarified various stages of sugar beet diseases as a continuum rather than discrete classes. Gradients of reflectance exist between healthy/asymptomatic and symptomatic tissue which may impede the classification between healthy and diseased leaf areas. The development of patterns in time and space, recorded by hyperspectral imaging may help to identify disease or stress influencing crops on the canopy level  and on the tissue level .
Given that the spectral patterns of healthy and diseased tissue are known, supervised classification was the choice to detect, identify, and quantify diseased tissue of sugar beet leaves. Since SAM classification is based on defined endmember spectra, the detection of leaf colonization prior to the occurrence of visible symptoms was not feasible by following this approach, but visible symptoms were classified with high accuracy. Benefits of the SAM algorithm for disease detection are insensitivity to heterogeneities of surface topography and illumination, because the angle between two vectors is invariant with respect to the length of the vectors . Leaf veins and differences in growth rates cause a characteristic undulated, grooved topography of sugar beet leaves depending on the genotype. Heterogeneities in reflectance intensity occur, as radiation is not reflected straightforward by these surfaces. Although classification accuracy of SAM was satisfying, it should be mentioned that this classification algorithm uses the average spectrum of each endmember class (e.g. healthy and different symptom peculiarities). The spectral variability within each endmember class, denoted as intra-class variability is not retained. Luc et al.  obtained a higher overall classification accuracy of Belgian coastline regions by modifying the common SAM to an optimized SAM preserving the intra-class variability. This approach may also resolve problems in disease classification, e.g. lower accuracy for early disease stages when only immature symptoms occur. Similar to the problems in the tomato - P. infestans system described by Zhang et al.  low disease levels of SBR resulted in lower accuracy of the SAM algorithm in this study.