Hyperspectral imaging for disease detection in seed potatoes Phenodays, October 28-30, 2015 Gerrit Polder, Pieter Blok, Jan Kamp - Wageningen UR Peter van der Vlugt - Kverneland Group Mechatronics

Outline

 Problem statement  History  Approach  Sensor techniques used  Hyperspectral Imaging  Results  Conclusion

Problem statement

 Virus and bacterial diseases are

one of the biggest problems in the cultivation of seed potatoes.

 Once found in the field, this lead to rejections of the tubers resulting in a big financial loss.

 Total area of seed potatoes in the Netherlands: ~ 40.000 ha

 Direct damage: 20-25M€ (6%)  Selection costs: 8–10M€  Total for a 40 ha company 30k€

How will the selection develop?

Goal

 Development of non-invasive fast sensors to detect virus and bacteria diseased plants in the open field.

● Test different sensor technologies. ● In the lab. ● On the field.

 Implement on autonomous robot, including

the handling of the diseased plants (remove, spray, …)

Experiments from previous years

 Greenhouse experiment, top view images.  Virus infected plants (cv Bintje): PVY NTN en PVY N-W

 Bacteria infected plants: Dickeya solani (2 levels)  Ground truth determined by bacteriologic tests  Results: ● Virus – 100% classification for full grown plants, less for earlier stages.

● Bacteria – 50% of the diseased plants were found, where the crop expert only found 30%.

Approach 2015

 Plants grown in pots in the field.  Virus (15 + 11 control) and Bacteria (80 + 40 control).  Weekly measurements with several techniques.  Compare with manual scores of crop expert. ● At what time step are symptoms seen by: ● Different techniques ● Crop expert

 Classify against ground truth: ● The amount of Virus infection. ● The amount of Bacteria found after harvesting.

Techniques

 Hyperspectral imaging  Thermal imaging (FLIR)  Chlorophyll Fluorescence imaging (Phenovation)  Force-A multiplex sensors. ● MX-330 (stilbenes) ● MX-375 (flavenols)

Virus detection in potatoes on efficiency of photosynthesis Presence of PAMV-Virus shows spots with low efficiency of photosynthesis

CropReporter

Colour (RGB)

Efficiency of photosynthesis

Thermal – water transport Erwinia E07

Healthy

Erwinia E07

Healthy

Erwinia E06

Erwinia E07

Hyperspectral imaging

 Research question: ● Is it possible to detect diseased potato plants in an early stage using the reflected light spectrum.

● Virus diseases (PVS, PVY, PAMV, PVV, PVA)

● Bacterial disease (Erwinia)

 Goal: ● Find most discriminating

wavelengths for implementation in a multispectral camera sensor.

Hyperspectral imaging lately attracted quite some attention

hyperspectral images are recorded by an imaging spectrograph (ImSpector V10e) placed between camera and lens A 1-D line falls on a prism-grating-prism which splits up the line in separate wavelengths

Hyperspectral top viewer

Hyperspectral recording 500 nm

700 nm

600 nm

800 nm

Results- Virus infected plants (top view)

Healthy

Infected

Hyperspectral side viewer

Differentiate between stem and leaves

Differentiate between stem and leaves

Based on a few example images, using a trained linear classifier Stem and Leaf pixels are selected for all images.

Analysis virus

 Random selection of leaf and stem pixels.  Gaussian classifier trained on infection yes/no at the time of planting:

 Leave one out cross validation per plant.  Majority voting determines class.

Preliminary results – Virus

 Symptoms: leaf

morphology (krinkel, deep veins)

Stem Leaves (score on symptoms)

Predi: True:

Healthy

Infected

Total

Healthy

7 (10)

3 (0)

10

Infected

1 (1)

13 (13)

14

Total

8 (11)

16 (13)

24

Predi: True:

Healthy

Infected

Total

Healthy

9 (10)

1 (0)

10

Infected

1 (1)

14 (14)

15

Total

10 (11)

15 (14)

25

Analysis bacteria

 Random selection of leaf and stem pixels.  Gaussian classifier trained on presence of D Solani at the stem base in 4 classes:

1. 2. 3. 4.

no infection low number of bacteria high number of bacteria very high number of bacteria

 Leave one out cross validation per plant.  Majority voting determines class.  Class 2-4 joined for classifying healthy/diseased.

Preliminary results – bacteria

 Symptoms:

wilted top leaves, dark top leaves, black stem.

Stem Leaves

(score on symptoms)

Predi: True:

Healthy

Infected

Total

Healthy

14 (15)

2 (1)

15

Infected

13 (7)

2 (8)

16

Total

27 (22)

4 (9)

31

Predi: True:

Healthy

Infected

Total

Healthy

15 (15)

1 (1)

15

Infected

15 (7)

0 (8)

16

Total

30 (22)

1 (9)

31

Preliminary conclusion

 Virus: ● Clear relation between spectral properties

measured from the side, with morphological features.

● Spectral reflection of leaves performs better than spectral reflection of stem.

 Bacteria: ● As long as there are no symptoms, almost no infected plants can be classified based on spectral reflection from the side.

● Even when symptoms are present, almost no relation.

Many thanks for your attention

Special thanks to: Pieter Kastelein

Jan van de Wolf Jan Willem van Leersum Vincent Jalink (Phenovation) Marc Pastor (Force-A)