Title: Spectral Weed Detection and Precise Spraying
1Spectral Weed Detectionand Precise Spraying
- Laboratory of AgroMachinery and Processing
- Els Vrindts, Dimitrios Moshou, Jan ReumersHerman
Ramon, Josse De Baerdemaeker
Research sponsored by IWT and the Belgian
Ministry of Small Trade and Agriculture
2Overview
- Spectral measurements of crops and weeds
- in laboratory
- in field
- Processing of spectral data with neural networks
- Precise spraying
3Optical detection of weeds
- Techniques
- red/NIR detectors (vegetation index)
- image processing (color, texture, shape)
- remote sensing of weed patches
- reflection in visible NIR light
- different detection possibilities, different
scales - Requirements for on-line weed detection
- fast accurate weed detection
- synchronized with treatment
4Spectral weed detection
- Factors affecting spectral plant signals
- leaf reflection, dependent on species and
environment, stress, disease - canopy measurement geometry
- light conditions
- detector sensitivity
5Spectral analysis of plant leavesin laboratory
Laboratory measurements
Diffuse Reflectance Spectroscopy of Crop and Weed
Leaves
6Diffuse Reflectance of a Leaf
Laboratory measurements
7Spectral Dataset
Laboratory measurements
8Reflectance of crop and weed leaves
Laboratory measurements
9Spectral analysis
Laboratory measurements
- stepwise selection of discriminant wavelengths
- multivariate discriminant analysis, based on
reflectance response at selected wavelengths
(dataset a) - assuming multivariate normal distribution
- quadratic discriminant rule
- classes with different covariance structure
- testing the discriminant function classification
of spectra from dataset b
10Spectral response of beet weeds
Laboratory measurements
11Spectral response of maize weeds
Laboratory measurements
12Spectral response of potato weeds
Laboratory measurements
13Classification results
Laboratory measurements
14Field measurement of crop and weeds
Field measurements
Signal path
Processingmethod
Variation inlight condition
Detector sensitivity
Measurement geometry
15Equipment for field measurement
Field measurements
spectrograph 10-bit CCD, digital
camera, computer, 12 V battery and transformer on
mobile platform
16Equipment - Spectrograph
Field measurements
both spatial and spectral information in images
17Image data
Field measurements
- maize, sugarbeet, 11 weeds
- 2 different days, different light conditions
- 755 x 484 pixels
18Spectral response of sensor
Field measurements
19Data processing
Field measurements
- spectral resolution 0.71 nm /pixel
- plant/soil discrimination with ratio NIR (745
nm) / red (682 nm) - data reduction by calculating average per 2.1 nm,
removing noisy ends - resulting spectra 484.8 - 814.6 nm range, 2.1 nm
step - independent datasets of maize, sugarbeet and weeds
20Spectral datasets
Field measurements
21Mean canopy reflections
Field measurements
22Canonical analysis of Sugarbeet - weeds
Field measurements
23Canonical analysis of Maize - weeds
Field measurements
24Discriminant analysis Sugarbeet
Field measurements
25Discriminant analysis Maize
Field measurements
26Graphic comparison datasets
Field measurements
27Graphic comparison datasets
Field measurements
28Graphic comparison datasets
Field measurements
29Discriminant analysis ratiosSugarbeet
Field measurements
30Discriminant analysis ratiosMaize
Field measurements
31Results
Field measurements
- only spectral info (485-815 nm)
- classification based on narrow bands in
discriminant functions - good results in similar light and crop conditions
- large decrease in performance for other light
conditions - using ratios of narrow bands
- improvement, but not sufficient
32Improving results
Field measurements
- influence of light conditions
- adaption of classification rule
- determining light condition and applying
appropriate calibration/LUT - spectral inputs that are less affected by
environment - measuring irradiance, calculating reflectance
- other classification methods
33Neural network for classification
Crop-weed classification
- Comparison of different NN techniques for
classification - Self-Organizing Map (SOM) neural network for
classification - used in a supervised way for classification
- neurons of the SOM are associated with local
models - achieves fast convergence and good
generalisation.
34Neural network for classification
Crop-weed classification
SOM
MLP
PNN
- ADVANTAGES
- Learns with reduced
- amounts of data
- Fast Learning
- Visualisation
- Retrainable
- DISADVANTAGES
- Discrete output
- ADVANTAGES
- Good extrapolation
- DISADVANTAGES
- Slow Learning
- Local minima
- Needs a lot of data
- ADVANTAGES
- Fast Learning
- Retrainable
- DISADVANTAGES
- Needs all training data
- during operation
- Needs a lot of data
35Comparison between methods
Crop-weed classification
MLP Multi-Layer Perceptron, PNN Probabilistic N
Network, SOM Self-Organizing Map, LVQ Learning
Vector Qantization, LLM Local Linear Mapping
Moshou et al., 1998, AgEng98, Oslo Moshou et al.,
2001, Computers and Electronics in Agriculture 31
(1) 5-16
36Crop-weed classification
Comparison between methods
MLP Multi-Layer Perceptron, PNN Probabilistic N
Network, SOM Self-Organizing Map, LVQ Learning
Vector Qantization, LLM Local Linear Mapping
37Crop-weed classification
Comparison between methods
MLP Multi-Layer Perceptron, PNN Probabilistic N
Network, SOM Self-Organizing Map, LVQ Learning
Vector Qantization, LLM Local Linear Mapping
38Conclusions on LLM SOM technique
Crop-weed classification
- The strongest point is the local representation
of the data accompanied by a local updating
algorithm - Local updating algorithms assure much faster
convergence than global updating algorithms (e.g.
backpropagation for MLPs) - Because of the topologically preserving
character of the SOM, the proposed classification
method can deal with missing or noisy data,
outperforming optimal classifiers (PNN) - The proposed method has been tested and gave
superior results compared to a variety
statistical and neural classifiers
39Precision spraying through controlled dose
application
Precision treatment
Unwanted variations in dose caused by horizontal
and vertical boom movements
40Active horizontal stabilisation of spray boom
Precision treatment
- Validation with ISO 5008 track
- movement of spray boom tip with and without
controller
41Vertical stabilisation of spray boom
Precision treatment
Slow-active system for slopes
Resulting boom movement
42On-line selective weed treatment
Precision treatment
Indoor test of on-line weed detection and
treatment
43Indoor test of on-line weed detection and
treatment
Precision treatment
- Sensor Spectral line camera
- Classification Probabilistic neural network
- Program in Labview with c-code
- Image acquisition frequence 10 images/sec,
travel speed 30cm/sec, segmentation with NDVI (
gt 0.3) - Off-line training of NN, On-line classification
- Decision to spray
- gt 20 weed pixels and gt 35 of vegetation is weed
- Spray boom with PWM nozzles and controller,
provided by Teejet Technologies
44Indoor test of on-line weed detection and
treatment
Precision treatment
- Color image and spectral image
45Indoor test - Results
Precision treatment
- Comparison of nozzle activation with weed
positions
46Indoor test - Results
Precision treatment
- separate weed classes (4) did not improve
crop-weed classification - Correct detection of nearly all weeds
- Only 6 redundant spraying of crop
- Up to 70 reduction of herbicide use
Experimental set up
camera
nozzle
weed