cuulee/satelliteimageclassification

Pixel based classification of satellite imagery - feature generation using Orfeo Toolbox, feature selection using Learning Vector Quantization, CLassification using Decision Tree, Neural Networks, Random Forests, KNN and Naive Bayes Classifier

SatelliteImageClassification

Pixel based classification of satellite imagery

• sample training and testing points generated using Point Sampling plugin in QGIS
• feature generation using Orfeo Toolbox
• feature selection using Learning Vector Quantization
• CLassification using Decision Tree, Neural Networks, Random Forests, KNN and Naive Bayes Classifier
• Ensemble classifier for Flood Inundation Mapping - classifies a pixel as water if 2 or more than 2 of the above classifiers classify a pixel as water
• Mode filter used to remove individually wrongly classified pixels
• Classification accuracy to measure goodness of each model

Outcomes of the best classifier (Multi Layer Perceptron) are as shown below:

To compile and run SatelliteClassification.java, you need weka.jar that you can download from the Weka website.

Compile code:  javac -cp weka.jar SatelliteClassification.java  
Run code: java -cp weka.jar:. SatelliteClassification  "trainingFile" "testingFile" "classifiername"

• order:

|clouds: white | |roads: yellow | |shadow: black | |urban: pink | |vegetation: green | |water: blue |

Original LANDSAT 8 Image during Flooding	Multi Layer Perceptron Classification
Ensemble Classifier: Water vs Everything, without filtering	Ensemble Classifier: Water vs Everything, after mode filtering

References:

• LANDSAT-8 imagery(http://earthexplorer.usgs.gov/)
• Image Preprocessing - QGIS (http://www.qgis.org/en/site/), ArcGIS (https://www.arcgis.com/features/index.html)
• Feature selection performed using Orfeo Toolbox (https://www.orfeo-toolbox.org/)
• Feature Selection: Vatsavai, Ranga Raju. "High-resolution urban image classification using extended features." In 2011 IEEE 11th International Conference on Data Mining Workshops, pp. 869-876. IEEE, 2011.