Figure 1.
Characteristic phenotypes of healthy and diseased barley leaves.
(A) Healthy barley leaf, (B) net blotch caused by Pyrenophora teres, (C) brown rust caused by Puccinia hordei, and (C) powdery mildew caused by Blumeria graminis hordei.
Figure 2.
Mapping disease progression in plants at massive scale.
Consecutive steps from hyperspectral imaging data of healthy and diseased barley leaves to interpretable summaries by metro maps in five consecutive steps. (Best viewed in color)
Figure 3.
Interpolated mean signatures and archetypal signatures for visible-near infrared (VNIR) and shortwave infrared (SWIR) wavelengths (measured 4–14 dai).
In the left column mean signatures of diseased barley plants before selecting disease archetypal signatures and in the right column mean archetypal signatures for η = 1 are illustrated. Archetypal signatures allow a better differentiation between different developing stages of the diseases. Moreover, they are in accordance to visually and manually extracted reflectance signatures during disease development. (Best viewed in color)
Figure 4.
Disease archetypal selections.
An example image showing diseased barley plants (RGB, first column) with powdery mildew (first row), net blotch (second row) and rust (third row) 14 dai. False color images present automatically determined diseased plant pixels based on disease archetypal signatures for VNIR and SWIR data (middle and right columns). The yellower/redder the color, the greather the difference of the pixel to a healthy plant. (Best viewed in color)
Figure 5.
Single sketches of hyperspectral dynamics of plant diseases for visible-near infrared (VNIR) wavelengths.
Each sketch consists of parts encoding major states during pathogenesis of the plant disease with similar weights. Thus, the shorter a part, the higher the impact of the corresponding period. (Best viewed in color)
Figure 6.
Collective disease progression via Metro Maps of hyperspectral dynamics of diseased plants for visible-near infrared (VNIR) (top) and short-wave infrared (SWIR) wavelengths (bottom).
Each disease track from hyperspectral images exhibits a specific route in the metro map, the direction and the dynamic steps are in correspondence to biophysical and biochemical processes during disease development. The beginning of all routes is at the same time point/train station (day of inoculation, gray circle). (Best viewed in color)
Figure 7.
Example in which way Simplex Volume Maximization iteratively determines basis vectors for interpretable matrix factorization.
Figure 8.
Example for multiple sketches on a synthetic data set.
The data set was generated by two moving Gaussians (green and red) where the brightness encodes time (the darker, the later) (left). The corresponding embedding using MDS (middle), the black dots were fixed on both dimensions whereas for the green and red dots only the x-axis coordinates were fixed (xi = ti, where t is the temporal information) and y-coordinates were unknown (learned by MDS). (right) The connection graph defining the weight matrix W between different states of N sequences over time, e.g. the moving Gaussians: wij = 1 if the two elements i and j belong to the same time slice (column) or to the same entity (raw or the same color), say moving Gaussian, but at consecutive time points.