improving classification accuracy using knowledge based approach ali a. alesheikh a. talebzadeh f....
TRANSCRIPT
Improving Classification Accuracy Using Knowledge
Based Approach
Ali A. Alesheikh
A. Talebzadeh
F. Sadeghi Naeeni
Image interpretation by computer vision
Traditional strategiesTraditional strategies Knowledge-basedKnowledge-based
Levels of processing and representation Theory and concepts of knowledge-based system
Various errors in remotely sensed image analysis Various errors in remotely sensed image analysis Techniques for knowledge representation Techniques for knowledge representation
use of external knowledge for image interpretation Use of prior probabilities in the decision rule Use of prior probabilities in the decision rule Use of other images as external knowledgeUse of other images as external knowledge
Implementation
Image information
Image Analysis
Computer Graphics
ArtificialIntelligence
ImageProcessing
- Traditional strategies
use very little knowledge about the domainuse very little knowledge about the domain
the most commonly used the most commonly used approachesapproaches in RS in RS
have various problemshave various problems
- Knowledge-based image interpretation
tends to use more external information in the inference processtends to use more external information in the inference process
use spectral information in the use spectral information in the imageimage
GIS Knowledge base
Matching goal achievement
inference
Symbolic description
Hypothesis database
SegmentationSegmentationFeature Feature
extractionextractionPre-processingPre-processing
Image data
Levels of Levels of representrepresent
ationation
highhigh
IntermediateIntermediate(STM)(STM)
Levels of Levels of processingprocessing
lowlow
High (LTM)High (LTM)
lowlow
-Various errors in remotely sensed image analysis During data acquisition process
-Various errors in remotely sensed image analysis During data acquisition process
Nature of data
Adjacent pixels have influence on each otherAdjacent pixels have influence on each other
-Various errors in remotely sensed image analysis During data acquisition process
Nature of data
Adjacent pixels have influence on each otherAdjacent pixels have influence on each other
Land cover types do not fit into multiples of rectangular spatial unitsLand cover types do not fit into multiples of rectangular spatial units
-Various errors in remotely sensed image analysis During data acquisition process
Nature of data
Adjacent pixels have influence on each otherAdjacent pixels have influence on each other
Land cover types do not fit into multiples of rectangular spatial unitsLand cover types do not fit into multiples of rectangular spatial units
Different surface materials may be distinguished by very Different surface materials may be distinguished by very subtle differences in their spectral patternssubtle differences in their spectral patterns
-Various errors in remotely sensed image analysis
During data acquisition process
Nature of data
During classification process
Adjacent pixels have influence on each otherAdjacent pixels have influence on each other
Land cover types do not fit into multiples of rectangular spatial unitsLand cover types do not fit into multiples of rectangular spatial units
Different surface materials may be distinguished by very Different surface materials may be distinguished by very subtle differences in their spectral patternssubtle differences in their spectral patterns
Types of knowledge
Knowledge-a priori domain dependent
Knowledge-a priori domain dependent
declarativedeclarative
heuristicheuristic
algorithmalgorithm
inheritableinheritable
non-inheritablenon-inheritable
optionaloptional
essentialessential
negativenegative
relationalrelational
proceduralprocedural
objectobject
• semantic network
IF <condition> THEN <action>
represents objects and relations between objects as a graph structure i.e. a set of nodes connected by labeled arcs
In a frame-based system the objects at each node in the network is defined by a collection of attributed, slots, and values of thoseattributes, called fillers. Each slot can have procedures attached to it
• production rules
• frames or schemas
Rule #1
IF a pixel feature is (92,99,91) THEN it is “W (Wheat)” or “BID (Barely)” or “SB (Sugar beet)” or “ALO (Alfalfa)”.
Rule #2 IF a region in Aster's NDVI map is lower than 0.15 e
THEN it's crop type will be W (Wheat) or BID (barely).Rule #3
IF last year's crop was MS THEN in the interest year the crop will be W (Wheat).
Example of each knowledge representation techniques
BID W
Last year's crop was MS
ALO SB MS
MF
MG Maximum probability in traditional classification (e.g. maximum likelihood
classification)
Value on Aster's NDVI map on August
<0.15
is a
is a is a is a is a is a is a
Example of each knowledge representation techniques
Frame “W,BID,SB,ALO” slot: they are: W(Wheat),BID(Barely),SB(Sugar beet),ALO(Alfalfa). procedure: if identification of them is desired then search pixels that have maximum probability in any traditional classification like maximum likelihood classification.End frameFrame “W,BID” slots : they are: W(Wheat), BID(Barely). criterion for reconnaissance: they are harvested on the middle of June. procedures: if recognition of W or BID between recognized W, BID, SB, ALO is desired then search areas on Aster's NDVI map which is lower than 0.15.End frameFrame “W” slots : is: W(Wheat), is generalization of: W17, W22, WAT, WTN, WP, WKU, WGP. criterion for reconnaissance: for using the soil in the best way to producing crops, crop calendar disciplines must be considered. procedures : if reconnaissance of W between recognized W, BID is desired then we can use crop calendar disciplines, e.g. search the areas that their last year's crop was MS(Maize Seed). End frame
Example of each knowledge representation techniques
1
Realthreshold
Estimated threshold
Real distribution of
class 1
A posteriori probability of class 2 given equal a prior
probability
Probability
Feature
A posteriori probability of class 1 given equal A prior
probability Real distribution of
class 2
- Using of prior probability in the decision rule (maximum likelihood approach)
1
P{w k ,Xi }
P{X}P {w k| Xi} =
P{wk ,Xi } = (Xi) P{wk }
-p/2|e-1/2(X-'^ (-1)X-)
P{w k | Xi , v j} = Xi P{w k , v j }
K
k=1Xi P{w k , v j }
1
- Using of prior probability in the decision rule (maximum likelihood approach)
1
P{w k ,Xi }
P{X}P {w k| Xi} =
P{wk ,Xi } = (Xi) P{wk }
-p/2|e-1/2(X-'^ (-1)X-)
P{w k | Xi , v j} = Xi P{w k , v j }
K
k=1Xi P{w k , v j }
- Using of other images as external knowledge
The other knowledge for interpretation can be the other image which is acquired in the other time or with the other sensor. The resolution and spectral bands of the other image can be different from initial one.
Study area
• Moghan plain located in Ardebil
Study area
• Moghan plain located in Ardebil
• About 300,000 tons of various crops produce annually in 18000 ha of irrigated farms.
Study area
• Moghan plain located in Ardebil
• About 300,000 tons of various crops produce annually in 18000 ha of irrigated farms.
Corp Acreage(ha) Yield
Wheat 7000 up to 6500 kg/ha
Barely 1500-2000 up to 5000 kg/ha
Sugar Beet 3000 more than 50tons/ha
Maize Seed 15000 more than 2500 kg/ha
Maize Grain 1500 more than 6500kg/ha
Maize Silage 800 more than 40tons/ha
Alfalfa 1500 about 12tons/ha
Forage crops 700 20-100tons/ha
Available DATA:
• Maps of study area in 1/50000 scale
(UTM coordinate system and in WGS84 ellipsoid)
Available DATA:
• Maps of study area in 1/50000 scale
(UTM coordinate system and in WGS84 ellipsoid) • Map of field boundaries
( production of polygonized fields)
Available DATA:
• Maps of study area in 1/50000 scale
(UTM coordinate system and in WGS84 ellipsoid) • Map of field boundaries
( production of polygonized fields)
• Data about crop type of each field
Available DATA:
• Maps of study area in 1/50000 scale
(UTM coordinate system and in WGS84 ellipsoid) • Map of field boundaries
( production of polygonized fields)
• Data about crop type of each field
• ETM+ image (color composite 354)
(was acquired on 2001-05-23)
GIS of GIS of Moghan Moghan
FieldsFields
Available DATA:
• Maps of study area in 1/50000 scale
(UTM coordinate system and in WGS84 ellipsoid) • Map of field boundaries
( production of polygonized fields)
• Data about crop type of each field
• ETM+ image (color composite 354)
(was acquired on2001-05-23)
• Aster image
(was acquired on August 2001-8-23)
GIS of GIS of Moghan Moghan
FieldsFields
Available DATA:
• Maps of study area in 1/50000 scale
(UTM coordinate system and in WGS84 ellipsoid)
Georeferenced
by map on 1/50000 scale
• Map of field boundaries
( production of polygonized fields)
• Data about crop type of each field
• ETM+ image (color composite 354)
(was acquired on 2001-05-23)
• Aster image
(was acquired on August 2001-08-23)
GIS of GIS of Moghan Moghan
FieldsFields
Experimental work
• Spectral-based :
Crop rotation patterns Times of planting and harvesting Field boundaries information Climate information
× Geographical information
Experimental work
• Spectral-based :
• Knowledge-based : Crop rotation patterns Times of planting and harvesting Field boundaries information Climate information
× Geographical information
Experimental work
• Spectral-based :
• Knowledge-based : Crop rotation patterns Times of planting and harvesting Field boundaries information Climate information
× Geographical information
Experimental work
• Spectral-based :
• Knowledge-based : Crop rotation patterns Times of planting and harvesting Field boundaries information Climate information
× Geographical information
× Financial information × Crop 'portfolio management' × Agricultural information × Advice centers
Spectral-based : rule matrices of every seven crop based on maximum likelihood
approach and equal prior probability
Spectral-based : rule matrices of every seven crop based on maximum likelihood
approach and equal prior probability
Spectral-based : rule matrices of every seven crop based on maximum likelihood
approach and equal prior probability
Overall accuracy of spectral-based classification = 53.2%.
Spectral-based : rule matrices of every seven crop based on maximum likelihood
approach and equal prior probability
Overall accuracy of spectral-based classification = 53.2%.
- Using of Crop Rotation Patterns :
TRANSITION MATRIX "1997-1998, 1998-1999"
ALO BID MF MG MS SB W
ALO 0.91 0 0 0 0 0 0.09
BID 0 0.38 0.22 0.14 0.09 0.17 0
MF 0 0.47 0 0 0 0.05 0.48
MG 0 0.09 0 0 0 0 0.91
MS 0 0 0 0 0 0 1
SB 0 0 0 0 0 0 1
W 0 0.17 0 0.07 0.32 0.38 0.06
• Transition matrix production
Knowledge-based classification :
TRANSITION MATRIX "1998-1999, 1999-2000"
ALO BID MF MG MS SB W
ALO 0.86 0 0 0 0 0 0.14
BID 0 0.33 0.21 0.2 0.07 0.19 0
MF 0 0.49 0 0 0.04 0 0.47
MG 0 0.18 0.01 0 0 0 0.81
MS 0 0 0 0 0 0 1
SB 0 0.02 0 0 0.03 0 0.95
W 0 0.17 0.05 0.03 0.43 0.23 0.09
- Using of Crop Rotation Patterns :
TRANSITION MATRIX "1997-1998, 1998-1999"
ALO BID MF MG MS SB W
ALO 0.91 0 0 0 0 0 0.09
BID 0 0.38 0.22 0.14 0.09 0.17 0
MF 0 0.47 0 0 0 0.05 0.48
MG 0 0.09 0 0 0 0 0.91
MS 0 0 0 0 0 0 1
SB 0 0 0 0 0 0 1
W 0 0.17 0 0.07 0.32 0.38 0.06
• Comparison between them Stable Dynamic System
Knowledge-based classification :
TRANSITION MATRIX "1998-1999, 1999-2000"
ALO BID MF MG MS SB W
ALO 0.86 0 0 0 0 0 0.14
BID 0 0.33 0.21 0.2 0.07 0.19 0
MF 0 0.49 0 0 0.04 0 0.47
MG 0 0.18 0.01 0 0 0 0.81
MS 0 0 0 0 0 0 1
SB 0 0.02 0 0 0.03 0 0.95
W 0 0.17 0.05 0.03 0.43 0.23 0.09
GIS Information extraction
Terrain object data (t-1)
Remote sensing data (t)
Application context
Updating
IF last year's crop = WheatTHEN current crop = Barely (17%), Maize feed (5%), Maize grain (3%),
Maize seed (43%), Sugar beet (23%), Wheat (9%).
Overall accuracy of maximum likelihood and estimated prior probability 66.7%.
Knowledge-based classification :
- Times of planting and harvesting
Wheat and Barely are harvested on the June
Using of NDVI produced from Aster image which was acquiredon 23 August 2001
> 0.15< 0.15
Knowledge-based classification :
- Times of planting and harvesting
Wheat and Barely are harvested on the June
Using of NDVI produced from Aster image which was acquiredon 23 August 2001
IF value of NDVI map is smaller than 0.15THEN crop type will be W(Wheat) or BID(barely)IF produced probability of W from the previous step is greater than probability of BIDTHEN crop type will be W(Wheat)
Overall accuracy of knowledge-based classification = 72.3 %.
Knowledge-based classification :
- -Field boundaries Field boundaries informationinformation
In each field one crop typeIn each field one crop type
Overall accuracy of knowledge-based classification = 88.7 %.
conclusion
This paper shows us that "traditional image analysis seems to be like a random walk in problem space" and by using any external knowledge, known way can be selected for receiving the goal.
Future works• Crop rotation was used in this thesis. Transition matrices were produced from two
successive years. They can be extracted from three, four or more successive years.
• Other data sources can be used as external knowledge, e.g. the other bands of aster image can help us for interpretation.
• Knowledge about local soil types and conditions could be used to help predict likely crops to be planted.
• We can use geographical information as an external knowledge. E.g. economical constraints affect likelihood of crops. For example, crops with a high transportation cost and low profit margin may become less probable the further away from a storage silo the field is.
• Financial information can help us for image interpretation. By this fact that, farmers also base their decisions about which crops to plant based on market potentials, aiming to maximize profitability. Information about expected crop prices and likely future demand could again assist in classification