OIL PALM LEAF NUTRIENT ESTIMATION USING OPTICAL SENSORS
1. 1
OIL PALM LEAF NUTRIENT
ESTIMATION USING OPTICAL
SENSORS
Chairman:
• Associate Professor Dr. Abdul Rashid B. Mohamed Shariff
Members:
• Professor Dr. Shattri Mansor
• Associate Professor Dr. Anuar Abd Rahim
• Associate Professor Dr. Ali Sharifi
By: Khosro Khorramnia, PhD student
Faculty of Engineering, Biological & Agricultural Engineering Dept.
Oct. 25, 2012
2. 2
Introduction
Current Methodology: Leaf Sampling and chemical
analysis
1% of the plantation : 14,800 trees in 10,000 Ha plantation
Time consuming and costing
Fully destructive
No potential for precision agriculture
We need to develop Precision Agriculture
techniques and technologies to:
Improve productivity (reduce costs and increase outputs)
Reduce environmental impacts
Overcome lack of farm workers
Increase job satisfaction
Attract highly educated people in ICT to agriculture careers
By: Khosro Khorramnia
3. Objectives
1. Determine the relationship between soil electrical
conductivity and soil nutrient availability in a typical
oil palm farm
2. Determine the relationship between different
vegetative indices such as the Normalized Deviation
Vegetation Index (NDVI) and oil palm nutrient
deficiency by conducting field experiments.
3. Providing essential knowledge to develop decision
support system for variable rate application of
fertilizers based on spectral reflectance and soil
electrical conductivity measures.
By: Khosro Khorramnia
4. 4
Materials and Methods:The Research was
done in two phases (1)
1. Phase one: The objectives of the first phase experiments
were;
A. Selecting the best types of different sensors suited
for oil palm foliar sensing.
B. Doing preliminary soil and leaf chemical analysis to
figure out any significant relationships between soil
chemical characteristics and oil palm leaf nutrient
status.
C. Developing prediction models to estimate leaf
Nitrogen (N), Phosphorous (P), Potassium (K),
Magnesium (Mg), Calcium (Ca) and Boron (B) status
using non-destructive sensing methods.
2. Phase two
By: Khosro Khorramnia
5. 5
Materials and Methods:The Research was
done in two phases (2)
1. Phase one
2. Phase two: The objectives of the second phase
experiments were;
A. Selecting the best types of different sensors suited
for oil palm foliar sensing.
B. Developing prediction models to estimate leaf
Nitrogen (N), Phosphorous (P), Potassium (K),
Magnesium (Mg), Calcium (Ca) and Boron (B) status
using non-destructive sensing methods.
By: Khosro Khorramnia
6. I II III IV
Specifications/sen
sor
Handheld
reflectance sensor
Chlorophyll
meter
Fluorescence sensor Spectroradiometer
Manufacturer NTech Industries Inc.,
CA, USA
Minolta
Sensing, Inc.
Osaka, Japan
Dynamax Inc. TX, USA ASD Inc., CO, USA
Model GreenSeeker
®
RT505 SPAD 502 Plus Multiplex
®
3 FieldSpect
®
3, Hi-Res
22 spectral features:
Yellow (YF), Red (RF),
Far Red (FRF)
fluorescence for ultra
violet (UV), Blue (B),
Green (G) and Red (R)
Light Source LED 2 LEDs LED (pulsed operation) , 4
excitation channels: UV,
Blue, Green and Red
Halogen bulb
Power 12 VDC Two AA-size
batteries (1.5V)
External Li-ion
rechargeable battery
12-18 VDC, 6.5 W
Visible (660 nm),
Near infra-red (770 nm)
Foot print/ Field of
view
1 × 60 cm 2 × 3 mm < 10 cm diameter and 50
sq cm
10 mm diameter
Operating Height 85 cm Clamp the
meter over
leafy tissue
10 cm Clamp the meter over
leafy tissue
Mount Handheld Handheld Handheld Portable
Spectral Features Sensing optical
density
difference at
two
wavelengths by
silicon
photodiode
Reflectance spectra (350-
2500 nm) with sampling
rate of 1.4 nm for 350-
1050 nm and of 2 nm for
1000-2500 nm
Operational
Wavebands
650 nm and 940
nm
UV(373), Blue(470),
Green(516), Red - orange
(635), YF(590), RF(685),
350 - 2500 nm
Types of sensors
By: Khosro Khorramnia
7. 7
Materials and Methods:
Study areas
1. UPM oil palm farm (two
replications): First phase
2. Sime Darby Plantation, Carey
Island, three different farms:
Second phase
By: Khosro Khorramnia
9. 9
Methodology:
Three datasets were used
Visible-near infrared reflectance dataset
Fluorescence dataset
GreenSeeker and SPAD 502 dataset
Data size consisted of 164 samples (trees)
Outliers detected (Grubb's Test or extreme studentized
deviate ) and removed form the data sets. Data
size reduced to 153. 70% as training data sets
and 30% as testing.
By: Khosro Khorramnia
10. 10
Methodology: Data mining and Model Making
1. Stepwise Regression Analysis to reduce the
size of data and model making.
2. Artificial Neural Network, using selected
predictors by Stepwise regression Analysis.
By: Khosro Khorramnia
11. 11
Results and discussion: Phase one
Phase one
1. There is no significant relationship between soil
chemical characteristics and leaf nutrient status.
2. GreenSeeker and SPAD 502 can not be useful to
estimate leaf nutrient status alone and other types
of sensors should be examined.
By: Khosro Khorramnia
15. Receiver operating characteristic (ROC)
ROC curves provide a comprehensive and visually
attractive way to summarize the accuracy of
predictions.
(ROC) analysis is an established method of
measuring diagnostic performance in medical
imaging studies.
Recently researchers have begun to report ROC curve
results for ANN classifiers.
By: Khosro Khorramnia
16. Terms of ROC
ONE method of specifying the performance of a
classifier is to note its true positive (TP) rate and false
positive (FP) rate for a data set. The TP rate is the
percentage of target samples that are correctly
classified as target samples.
The TP rate is commonly referred to as “sensitivity,”
and (1—FP rate) is called “specificity.”
By: Khosro Khorramnia
17. True disease state vs. Test result
Power 1 - β;
sensitivity = TP
X
Type II error (False
-) β
Disease
(D = 1)
X
Type I error (False
+) α
specificity = (1-FP)
No disease
(D = 0)
rejectednot rejected
Disease
Test
By: Khosro Khorramnia
29. 29
Artificial Neural Networks
1. Probabilistic Neural Networks (PNN) and
2. Multi-Layer Feed forward Networks (MLF)
Numeric predictions can be performed using MLF
networks, as well as Generalized Regression Neural
Networks (GRNN), which are closely related to PNN
networks.
By: Khosro Khorramnia
30. Artificial Neural Networks
Critical nutrient concentration ranges of frond #17 in oil
palm trees (Fairhurst & Hardter, 2003) – check ROC
By: Khosro Khorramnia
32. 32
Artificial Neural Networks: Conclusion
This model was developed for oil palm trees between 5-6 years old.
•Soil and climatic zoning: The whole oil palm area should be categorized into its soil and climatic
subdivisions.
•Soil and leaf sampling should be done simultaneously with leaf scanning using different sensors
involved in this study. It is assumed that 160-180 trees could be sufficient out of 10,000 Ha. In
conventional system: 148×100 = 14,800 trees are proposed for leaf sampling.
•It is better to choose three different types of trees with deficiency, optimum, and excess amount of leaf
nitrogen content.
•Using ANN or multi-regression analysis (depends on final experiment and research verification) an
appropriate model needs to be developed that applies to all plantation areas.
•This procedure should be applied every year to develop a robust and reliable model, so that leaf
sampling can be eliminated from the procedure when highest accuracy of predictions achieved. As the
tropical regions mostly have stable climatic conditions, the period of time to develop the model is
basically dependent to land management. The land (soil) and slop are basic sources of variations and
determination of different zones would reduce this period to develop the optimized model. The age and
variety of oil palm trees are other criteria for zoning.
•Fertilizers could be recommended based on predicted leaf nutrient contents, calculated from the
developed model.
By: Khosro Khorramnia
33. Probabilistic and General
Regression Neural Networks
Probabilistic (PNN) and General Regression Neural
Networks (GRNN) have similar architectures, but there is a
fundamental difference: Probabilistic networks perform classification
where the target variable is categorical, whereas general regression
neural networks perform regression where the target variable is
continuous. PNN and GRNN networks have advantages and
disadvantages compared to Multilayer Perceptron networks:
It is usually much faster to train a PNN/GRNN network than a
multilayer perceptron network.
PNN/GRNN networks often are more accurate than multilayer
perceptron networks.
PNN/GRNN networks are relatively insensitive to outliers (wild
points).
PNN networks generate accurate predicted target probability scores.
PNN networks approach Bayes optimal classification.
PNN/GRNN networks are slower than multilayer perceptron
networks at classifying new cases.
PNN/GRNN networks require more memory space to store the
model.
By:KhosroKhorramnia
Expectations after research done
Increase our knowledge about oil palm
Develop new models for fertilizer requirement assessment
Develop DSS for leaf Nitrogen sensing
Formulate a new system for VRA of Fertilizers
Interface Development
Suggestions for future researches