vegetation indices
The following results were achieved in cooperation with Till Rumpf, Institute of Geodesy and Geoinformation, Department of Geoinformation, University Bonn, using Support Vector Machines (SVMs) with SVIs as features for detection and classification of the diseases.
4.4. Detection and classification of plant diseases with Support Vector Machines based on spectral vegetation indices
4.4.1 Dichotomous classification between healthy and diseased sugar beet leaves
In a first dichotomous approach Support Vector Machines were used for the differentiation between two classes, non-inoculated, healthy leaves and leaves, inoculated with one of the three leaf pathogens, respectively. Eight SVIs (NDVI, SR, SIPI, PSSRa, PSSRb, ARI, REP, mCAI) and SPAD-values were used as features for classification. The results showed that the specificity of classifica- tion was always above its sensitivity. Accuracy ranged from 93% to almost 97% (Tab. 4.6).
The classification accuracy increased with increasing disease severity (Fig.4.17). Differences in the number of leaves per disease give additional information on the reliability of classification results. With only 1% to 2% diseased leaf area, the classification accuracy was about 65% for all diseases. The accuracy of dif- ferentiating between healthy leaves and leaves with Cercospora leaf spot symp- toms rapidly increased with 3% to 5% disease severity. When more than 10% of the leaf area was covered by leaf spots, the classification accuracy reached 100% (Fig. 4.17). The accuracy of detecting symptoms of sugar beet rust and powdery mildew was lower at low disease severities. For sugar beet rust classifi- cation accuracy reached 95% when disease severity was above 6%. Leaves with powdery mildew could be differentiated from healthy leaves with an accuracy of about 95% when 10% to 15% of the leaf area was diseased (Fig. 4.17).
Table 4.6: Classification result of the dichotomous classification between healthy and diseased sugar beet leaves based on spectral vegetation indices using Support Vector Machines.
Leaf disease Accuracy [%] Specificity [%] Sensitivity [%]
Cercospora leaf spot 96.68 97.84 95.45
Sugar beet rust 96.20 97.14 95.14
4. RESULTS
Figure 4.17:Classification result of non-inoculated sugar beet leaves and sugar beet leaves with three diseases depending on disease severity (numbers within bars denote class size).
4.4.2 Multi-class classification among healthy leaves and leaves with specific disease symptoms
Tab. 4.7 summarises the results of the model learned which classified healthy sugar beet leaves and leaves diseased with Cercospora leaf spot, sugar beet rust, and powdery mildew, respectively (multi-class classification). The overall classification accuracy was better than 88%, differences between the classes were low. The class recall of each class ranged between 84% and > 92%. The class of healthy leaves was classified best. Classification difficulties occurred in separating between sugar beet rust and Cercospora leaf spot and also in the classification between powdery mildew and healthy sugar beet leaves.
4.4.3 Classification of healthy leaves and leaves inoculated with fun- gal pathogens at early stages of pathogenesis
For the differentiation between healthy sugar beet leaves and leaves inoculated with one of the pathogens before specific disease symptoms became visible, SVI data were used starting 3 dai. First symptoms of Cercospora leaf spot
4.4. Detection and classification of plant diseases with Support Vector Machines based on spectral vegetation indices
Table 4.7: Classification results of the multi-class classification between healthy and diseased sugar beet leaves based on spectral vegetation indices using Support Vector Machines.
Ground truth
Cercospora Sugar beet Powdery Class
Prediction Healthy
leaf spot rust mildew precision
Healthy 942 32 47 69 86.42%
Cercospora leaf spot 12 748 61 13 89.69%
Sugar beet rust 20 88 622 14 83.60%
Powdery mildew 46 12 10 834 92.46%
Class recall 92.35% 85.00% 84.05% 89.68% 88.12%
appeared 6 dai, rust 8 dai, and powdery mildew 5 dai. Leaves inoculated with
C. beticola were correctly classified by SVMs with an accuracy range from 65% to 80%, even before symptoms became visible (Fig. 4.18 A). When specific symptoms occurred 6 dai, the classification accuracy increased until 12 dai when it converged at 100%. Throughout the 21 days of the experiment, the classification accuracy of the automatic procedure was consistent to visually classified Cercospora leaf spot-infected leaves. The classification accuracy of healthy leaves was almost 87% 3 dai and reached > 95% 8 dai and later.
Although first symptoms of sugar beet rust appeared only 8 dai, a classification accuracy of 90% for U. betae-infected leaves was reached 3 to 5 dai (Fig.4.18B). One day before first rust symptoms became visible, the sensitivity decreased to about 71% and increased again 15 dai to > 98%. The results of the automatic classification were inferior to visual ratings only between 12 and 14 dai. Healthy sugar beet leaves were classified with an accuracy of 72% at early stages of leaf colonization by U. betae, but from 10 dai the classification accuracy was always > 95%.
Already 3 dai the classification accuracy of powdery mildew was > 80% and increased to almost 92% 5 dai (Fig. 4.18 C). After the appearance of first visible colonies the classification rate decreased < 70% 6 dai, and subsequently increased again from 10 to 20 dai > 95%. In contrast to the results for the leaves colonized by the other pathogens the visual classification of E. betae-infected
4. RESULTS
leaves was superior (13% to 23%) to the automatic classification procedure from 6 to 9 dai. In the beginning of the experiment, the classification rate of healthy leaves was > 91%; in the third week, however, it decreased to 77% 20 dai.
Figure 4.18: Effect of incubation time on the Support Vector Machines classification based on SVIs between healthy sugar beet leaves inoculated with Cercospora beticola (A), Uromyces betae (B), and