Comprehensive Hyperspectral Analysis for Indonesian Rice Agricultural Needs involving Climate and Social Dynamics

Page 1

Sidik Mulyono,

M. Ivan Fanany

Faculty of Computer ScienceUniversitas Indonesia

Indonesia Agency for The Assessment and Application of Technology

Comprehensive Hyperspectral

Analysis for Indonesian Rice

Agricultural Needs involving

Climate and Social Dynamics

Page 2

OUTLINE

1. Introduction:

Indonesia at a glance

Background

2. Our research

GA-NSPCR for feature

selection and yield

prediction using

hyperspectral

Balanced branches SVM

for growth stages

classification

An ensemble incremental

approach ELM for growth

stages classification

3. Related Studies

4. Remaining Problems &

Discussion

Page 3

1. INTRODUCTION

Page 4

a. Indonesia at the glance

Indonesia Archipelago

Rice field area: 8.1 million hectares

Population: 240 million

National rice need: 32 million T/year

Page 5

b. Background

Paddy field monitoring in Indonesia, still uses direct human observation and based on statistical calculation

Often causes irregularities, since the results tend to be either excessived or low-estimated

More comprehensive prediction system for paddy fields are really needed

To support the National Food Security Program

Remote Sensing

Application

Page 6

2. OUR RESEARCH

Page 7

1. GA-NSPCR for features selection and paddy yield

prediction model using hyperspectral remote sensing

data (International Geo-science and Remote Sensing

Symposium - IGARSS 2012, Munich – Germany)

SAMPLING AREA: Karawang, West Java Indonesia

(06 15 46.24S, 107 25 05.1E)

Hyperspectral HyMap Image

Page 8



data

Original HyMap (124 bands) data

samples correspond to 55 Yield data

samples

Page 99



data

Page 1010



data

GA-NSPCR

(*) Haibo Yao, Lei Tian, A Genetic-Algorithm-Based Selective Principal Component Analysis (GA-SPCA) Method

for High-Dimensional Data Feature Extraction, IEEE Transactions on Geoscience and Remote Sensing, vol. 41,

no. 6, June 2003

Page 11

Statistical Validation

Overall accuracy

= 84.96%

Page 12

The composited Images from 10

lines of HyMap Hyperspectral

Airborne Campaign

(Karawang 2011)

Paddy yield distribution map

with GA-NSPCRTotal production: 186,898 Ton

Harvested Area: 30,966 Ha

Yield average: 6.00 Ton/Ha



data

Page 13

Why MODIS Images:

• Direct broadcast capability

• Free of charge

• Almost covers all of the

earth surface

• Acquire each 4-5 days at

the same place

A sample of MODIS surface

reflectance product (MOD 09) with

500 meter spatial resolution

For Karawang District, June 15,

2012

2. BB-SVM for paddy growth stages classification using

MODIS remote sensing images (International Conference

on Advanced Computer Science and Information System -

ICACSIS 2012)

Page 14

HyMap Hyperspectral

(4.4 m spatial res

124 band spectral res)

MODIS Multispectral

(1.000 m spatial res

7 band spctral res)

Hyperspectral based

MODIS integration model


MODIS remote sensing images

Yield

prediction

Growth

stages

Page 15

Collecting MODIS synthetic spectral data samples from original HyMap for Yield Prediction Model

Original HyMap (124

bands) data samples

correspond to 55 Yield

data samples

MODIS synthetic data

samples (band: 1, 2, 4, 5, 6, 7)



Page 16

Model calibration for Yield Predicion Using NSPCR

The Best Model



Page 17

Vegetative stage

Reproductive stage

Ripening stage

0 – 30 DAT 31 – 45 DAT 46 – 55 DAT

56 – 72 DAT 73 – 81 DAT 82 – 90 DAT

91 – 100 DAT 101 – 111 DAT 112 – 120 DAT

Paddy growth stages terminology defined by IRRI



DAT: Day after transplanting

Page 18

Spectral knowledge approach from Hyperspectral for paddy growth stages

Vegetative Stage Reproductive Stage

Ripening Stage



Page 19

MODIS spectral data collection for paddy growth stages

NoClass of growth

stagesNumber of samples

1 veg1 103

2 veg2 110

3 veg3 120

4 rep1 120

5 rep2 120

6 rep3 117

7 rip1 120

8 rip2 120

9 rip3 120

10 soil 120

11 cloud 85

Total samples 1,255

Number of MODIS bands 6



Page 20

Support Vector Machines for multi-class with:

• Radial basis function (RBF) kernel trick

Software:

• LIBSVM introduced by Chih-Wei Hsu et al

• Interactive Data Language (IDL) program



Page 21

SVM01

Cloud SVM02

Rep2 SVM03

Rep1 SVM04

Rep3 SVM05

Veg1 SVM06

Rip3 SVM07

Rip2 SVM08

Veg2 SVM09

Soil SVM10

Veg3 Rip1One against all

(OAA)-SVM approach

SVM01

SVM05

SVM09

SVM10

Rip3 Veg1

Rep1

SVM06

SVM07

SVM08

Soil Veg3

Veg2

Rip2

SVM02

SVM03

SVM04

Rip1 Rep2

Rep3

Cloud

Balanced branches

(BB)-SVM approach

Multi-class strategy



Page 22

C dan γ calibration for each classifier using grid-search

SVM01 SVM02 SVM03

SVM04 SVM05 SVM06

SVM07 SVM08 SVM09

SVM10



Page 23

Labeled two-class for OAA-SVM Labeled two-class for BB-SVM

No.SVMs

Classifier

Labels

(-1) (+1)

1 SVM01 cloud

soil, veg1, veg2, veg3,

rep1, rep2, rep3, rip1,

rip2, and rip3

2 SVM02 rep2

soil, veg1, veg2, veg3,

rep1, rep3, rip1, rip2,

and rip3

3 SVM03 rep1soil, veg1, veg2, veg3,

rep3, rip1, rip2, and rip3

4 SVM04 rep3soil, veg1, veg2, veg3,

rip1, rip2, and rip3

5 SVM05 veg1soil, veg2, veg3, rip1,

rip2, and rip3

6 SVM06 rip3soil, veg2, veg3, rip1,

and rip2

7 SVM07 rip2soil, veg2, veg3, and

rip1

8 SVM08 veg2 soil, veg3, and rip1

9 SVM09 soil veg3 and rip1

10 SVM10 veg3 rip1

No.SVMs

Classifier

Labels

(-1) (+1)

1 SVM01

soil, veg1,

veg2, veg3,

rep1, rip2,

rip3

rep2, rep3,

rip1, cloud

2 SVM02rep2, rep3,

rip1cloud

3 SVM03 rep2, rip1 rep3

4 SVM04 rip1 rep2

5 SVM05veg1, rep1,

rip3

veg2, veg3,

rip2, soil

6 SVM06veg2, veg3,

soilrip2

7 SVM07 veg2 veg3, soil

8 SVM08 soil veg3

9 SVM09 veg1, rip3 rep1

10 SVM10 veg1 rip3



Page 24

Calibration model for Individual (two-class) SVM classifier

ClassifierOAA BB-OAA

Train Test Val Train Test Val

SVM01 0.926 0.962 0.931 0.981 0.974 0.980

SVM02 0.849 0.884 0.853 0.959 1.000 0.964

SVM03 0.851 0.879 0.854 0.744 0.864 0.759

SVM04 0.842 0.871 0.845 0.957 0.967 0.958

SVM05 0.767 0.871 0.780 0.838 0.891 0.845

SVM06 0.948 0.966 0.951 0.934 0.949 0.936

SVM07 0.785 0.836 0.792 0.915 0.864 0.909

SVM08 0.883 0.931 0.889 0.662 0.700 0.667

SVM09 0.727 0.800 0.736 0.930 0.953 0.924

SVM10 0.767 0.767 0.767 0.713 0.750 0.717



Page 25

Validation result using multi-class SVM classifierClass/ Growth

Stages

Classification Accuracy

OAA-SVM BB-SVM

Cloud 0.318 0.718

Soil 0.217 0.467

Veg1 0.544 0.631

Veg2 0.055 0.627

Veg3 0.083 0.683

Rep1 0.067 0.842

Rep2 0.383 0.933

Rep3 0.068 0.795

Rip1 0.558 0.858

Rip2 0.275 0.808

Rip3 0.067 0.833

Total Accuracy 0.239 0.745



Page 26

Masking image by using standard paddy field Map

Original MODIS Image (1,000m)Standard paddy field mapStandard paddy field image



Page 27

IDL Program for paddy prediction



Page 28


2012-03-29



Page 29




2012-03-29

Page 30

Paddy planting & harvesting periods

Source : Sumarno, Paddy planting periods based management for national rice

production, Sinar Tani No. 3136, Year XXXVI



Planting Periods Code Month Harvesting Periods Code Month

1. Rainy season T1 Nov Main harvesting P1 Feb

T2 Dec P2 Mar

T3 Jan P3 Apr

T4 Feb P4 May

T5 Mar P5 Jun

2. Middle season T6 Apr Middle harvesting P6 Jul

T7 May P7 Aug

T8 Jun P8 Sep

T9 Jul P9 Oct

3. Drought season T10 Aug Drought harvesting P10 Nov

T11 Sep P11 Dec

T12 Oct P12 Jan

Page 31

-

2.00

4.00

6.00

8.00

10.00

12.00

14.00

16.00

18.00

1/2

9/2

01

2

2/2

9/2

01

2

3/3

1/2

01

2

4/3

0/2

012

5/3

1/2

01

2

6/3

0/2

01

2

7/3

1/2

01

2

8/3

1/2

01

2

9/3

0/2

01

2

10

/31

/20

12

11

/30

/20

12

Pro

du

cti

on

(in

milli

on

to

ns)

Date

Prediction for rice production in Java Island

Ha

rve

sti

ng

in

mid

dle

se

as

on

Ha

rve

sti

ng

in

ra

iny

se

as

on

Ha

rve

sti

ng

in

dro

ug

ht

se

as

on

13,56 M tons 11,65 M tons



Page 32

Growth-stages monitoring in time series

(In case of West Java)

Page 33

Field Campaign

Location:

Karawang and Indramayu District of West Java

(from 25 June 2012 to 31 July 2012)

50

0 m

500 m

3. An Ensemble Incremental Approach of Extreme Learning

Machine (ELM) For Paddy Growth Stages Classification

Using MODIS Remote Sensing Images (International

Conference on Advanced Computer Science and

Information System - ICACSIS 2013)

Page 34

Spectral Library

#1 Early vegetative stage #2 Late vegetative stage #3 Generative stage

#4 Harvesting stage #5 Bareland stage #6 Ploughing stage



Using MODIS Remote Sensing Images

Page 35

An ensemble method

Methods for constructing ensemble (Thomas G. Diettrich):

1. Bayesian voting : enumerating the hypotheses

2. Manipulating the training data: generating multiple

hypotheses

3. Manipulating the input features

4. Manipulating the output targets

5. Injecting randomness




Page 36

Basic Extreme Learning Machine (ELM)

• Originally proposed by Huang, is a simple

learning algorithm for single-hidden Layer Feed-

forward Neural network (SLFN).

• This method is based on the Moore-Penrose

pseudo inverse providing the minimum Least-

Squares solution of general linear systems

• The output of a SLFN with L hidden nodes is

mathematically expressed as follows:(Moore-Penrose pseudo inverse)




Error Minimized (EM-ELM):

• In contrast, EM-ELM is a simple and effective

approach to automatically determine the

number of hidden nodes in generalized SLFNs

[Guorui Feng].

• This approach can add random hidden nodes

to SLFNs one by one or group by group, during

the growth of the networks

Page 37




The Algorithms for an ensemble

incremental approach ELM

Page 38Basic ELM with RBF EM-ELM with RBF




Basic ELM with Additive EM-ELM with Additive

Page 39

Accuracy improvement

Method

Conventional method of

ELM

An ensemble incremental

approach of ELM

Kappa Number of

nodes

Kappa Number of

nodes

B-ELM Additive 0.775 13 0.883 27B-ELM RBF 0.818 19 0.939 40EM-ELM Additive 0.716 42 0.853 46EM-ELM RBF 0.825 45 0.867 34




Page 40

Class B-ELM-ADD B-ELM-RBF EM-ELM-ADD EM-ELM-RBF

Early vegetative 27.89% 29.02% 25.69% 26.66%

Late vegetative 57.32% 49.27% 62.55% 64.06%

Generative 0.92% 15.49% 0.00% 0.20%

Harvesting 8.32% 2.83% 5.46% 2.07%

Bareland 0.54% 0.09% 2.58% 2.65%

Ploughing 5.01% 3.30% 3.71% 4.36%

Cloud 0 0 0 0

Computational cost 3.623 6.476 4.492 5.625

Date: 2012/03/29 (Planting season)

B-ELM-ADD B-ELM-RBF EM-ELM-ADD EM-ELM-RBF




Page 41

Class B-ELM-ADD B-ELM-RBF EM-ELM-ADD EM-ELM-RBF

Early vegetative 0.49% 25.11% 0.12% 9.07%

Late vegetative 18.47% 19.25% 22.25% 16.76%

Generative 26.00% 17.79% 23.11% 23.58%

Harvesting 18.64% 21.43% 17.72% 17.91%

Bareland 1.13% 0.73% 2.27% 1.64%

Ploughing 35.28% 15.71% 34.53% 31.04%

Cloud 0.00% 0.00% 0.00% 0.00%

Computational cost 3.581 6.198 4.387 5.474

Date: 2012/08/03 (After harvested season)

B-ELM-ADD B-ELM-RBF EM-ELM-ADD EM-ELM-RBF




Page 42

3. RELATED STUDIES

Page 43

compa

1. Modeling Paddy Rice Yield Using MODIS

Data, Dailiang, et al, Agriculture and

forest meteorology, Elsevier 2014

2. Hyperspectral Versus Multispectral Crop

Productivity Modeling and Type

Discrimination For The HyspIRI mission,

Mariotto, et al, Remote Sensing of

Environment, Elsevier 2013

Page 44

Modeling Paddy Rice Yield Using MODIS Data,

Dailyang, et al, 2013

Dailiang Method

Input:

8-day PsnNet (MOD17A2 )

8-day LAI (MOD15A2 )

Harvest indices

Study area: Irrigated paddy field

Method: Linear regression

Spatial res: 1,000 m

Number of sample sites : 191

Yield accuracy: 0.5 t/ha

Our Method

Input:

Daily spectral domain (MOD09)

Yield


Method: Non linear regression

(Machine learning)

Spatial res: 500 m

Number of sample sites: 55

Yield accuracy: 0.68 t/ha

Page 45

Hyperspectral Versus Multispectral Crop Productivity

Modeling Type Discrimination For The HyspIRI mission,

Mariotto, et al, 2013

Mariotto Method

Input:

Hyperspectral (Hyperion & Field

Spectrometer)

Multispectral: Landsat-7 TM Plus,

(ETM+), Advanced Land Imager

(ALI), Indian Remote Sensing (IRS),

IKONOS, and QuickBird.

Biomass & yield of 5 crops


Method: Stepwise MLR& PCA for

band selection

Spatial res: 2.4 – 30 m

Number of sample: 1232

Yield accuracy: -

Classification accuracy: 45.7– 93.7%

Our Method

Input:

Hyperspectral: HyMap airborne

Multispectral: MODIS (MOD09)

Yield of paddy


Method: Machine learning (GA-

NSPCR, BB-SVM, and ELM)

Spatial res: 500 m

Number of sample: 1255

Yield accuracy: 82.1%

Classification accuracy: 85.3 – 93.9%

Page 46

4. REMAINING PROBLEMS &

DISCUSSION

Page 47

Rice issues in Indonesia

Page 48

THANK YOUQ & A

Comprehensive Hyperspectral Analysis for Indonesian Rice Agricultural Needs involving Climate and Social Dynamics

Documents