mapping winter-flooded rice field using a hybrid machine
TRANSCRIPT
Mapping Winter-Flooded Rice Field Using A Hybrid Machine Learning Algorithm
and Rule-Based Expert System
Yiwen Sun, Tiejun Wang Andrew K. Skidmore, Qi Wang, Baoping Qing
WHY MAPPING WINTER-FLOODED RICE FIELDS?
Winter-flooded rice fields are artificial wetland ecosystems within a traditional agricultural system.
Rice fields with flooded fallow in winter
WHY MAPPING WINTER-FLOODED RICE FIELDS?
Flooded rice fields maintain anequivalent functions as naturalwetlands and have been provento be of great value for waterbirdsin winter.
Main foraging habitats ofthe Crested Ibis
WHY MAPPING WINTER-FLOODED RICE FIELDS?
Identifying the distribution of winter-flooded rice fields is important for grain yield estimation and biodiversity conservation.
CHALLENGES
Difficulties in quantifying the winter-flooded rice fields using traditional ground survey method at broad scale
Remote Sensing
CHALLENGES
Identification of the highly dynamic, small and scattered patches of winter-flooded rice fields in mountainous area appears challenging.
Flooded rice fields may be spectrally confused with other land cover categories as lakes or riverbanks.
POTENTIAL SOLUTION
Combining conventional image classification methods with GIS expert systems for vegetation mapping may yield more accurate maps.
Paddy rice cultivation in mountains often occurs along the narrow river valleys with typical terrain features due to the limit of temperature and water availability.
STUDY OBJECTIVE
To test whether integrating topographic data into winter-flooded rice field mapping using a Bayesian expert system increases the mapping accuracy
STUDY AREA
Area: 3980 km2
DATA USED FOR CLASSIFICATION
Satellite images- 30-m Landsat TM images (winter and summer)
Ancillary topographic data - Elevation- Terrain position
Field data- Ground-truth data of seven land cover types(i.e. winter-flooded rice field, winter-dry rice field, rain-fed field, open water, forest, shrub/grass and others)
METHODS
Image classification- Support Vector Machine (SVM) classifier
- Hybrid classifier: SVM + Expert system
METHODS
|∑ | ; 1, 2, … ,
Key mechanism of the expert system
Bayes’ theorem
prior probability from rule images
conditional probabilityexpert knowledge
METHODSItem of evidence Winter-flooded rice field …
Elev
atio
n (m
)< 500 0.05 …
500-600 0.2 …
600-800 0.3 …
800-1000 0.05 …
1000-1300 0.08 …
>1300 0 …
Terr
ain
posi
tion Gully 0.4 …
Lower mid-slope 0.2 …
Mid-slope 0.05 …
Upper mid-slope 0 …
Ridge 0 …
METHODS
|∑ | ; 1, 2, … ,
Key mechanism of the expert system
Bayes’ theorem
prior probability from rule images
posterior probability to relabel pixels
conditional probabilityexpert knowledge
METHODS
Accuracy assessment- Overall accuracy, Kappa, producer’s and user’s accuracy
Classifier performance comparison- McNemar test for related samples
RESULTS
Land cover map produced by the SVM classifier
RESULTS
Land cover map produced by the hybrid classifier
RESULTS
SVM
Hybrid
88.1
89.3
71.8
90.4
Accuracy of winter‐flooded rice fieldUser Acc.(%) Prod. Acc.(%)
RESULTS
65758595
Producer's accuracy (%)
65758595
User's accuracy (%)
RESULTS
Overall accuracy Kappa
81.50%0.78
90.80% 0.89
Accuracy of land cover map
SVM Hybrid
McNemar Test χ2 = 34.03, P < 0.0001
CONCLUSIONS
The terrain data and expert knowledge used in the Bayesian expert system provide additional information that complements the spectral information contained in the satellite images.
The hybrid approach combining machine learning algorithm with a GIS expert system has a potential to further improve the land cover mapping accuracy when used with time series imagery.
THANK YOU