Some Topics in Remote Sensing Image Classification Yu Lu
Outline Introduction Relevance in spatial domain Relevance in spectral domain Relevance among multiple features
Outline Introduction Relevance in spatial domain Relevance in spectral domain Relevance among multiple features
Introduction Remote Sensing Image
Introduction Remote Sensing Image Multispectral image 4-7 bands TM10.45~0.52μm 蓝绿波段 TM20.52~0.60μm 绿红波段 TM30.63~0.69μm 红波段 TM40.76~0.90μm 近红外波段 TM51.55~1.75μm 近红外波段 TM610.4~12.5μm 热红外波段 TM72.08~2.35μm 近红外波段 Hyperspectral image Several hundreds of bands
Introduction Remote Sensing Image Classification Pixel labeling Semantic image segmentation Object class segmentation Standard data set One image with some pixels labeled, instead of a image database including multiple images
Introduction Indian Pines 92AV3C 0.4 m~2.5 m, 220 bands, 17 classes, 145*145 Background, Alfalfa corn-notill, corn-min grass/pasture, grass/trees, grass/pasutre-mowed, Hay-windrowed, oat, wheat, woods, soybeans-notill, soybeans-min, soybean-clean, Bldg-Grass-Tree-Drives, stone-steel towers
Introduction Indian Pines 92AV3C band 50 band 100 band 50 band 150 band 200 band 220
Introduction Flight line C1 0.4 m~1.0 m, 12 bands 10 classes, 949*220 Alfalfa, Br Soil, Corn, Oats, Red Cl, Rye, Soybeans, Water, Wheat, Wheat-2
Introduction Flight line C1 band1 band1 band3 band3 b a n d 12
Outline Introduction Relevance in spatial domain Relevance in spectral domain Relevance among multiple features
Relevance in spatial domain How to capture spatial relevance Features to capture spatial relevance Filtered features: gabor Statistical features: lbp sift
Relevance in spatial domain How to capture spatial relevance CRF
Relevance in spatial domain Classifier to capture spatial relevance Standard SVM [1] “A Spatial–Contextual Support Vector Machine for Remotely Sensed Image Classification” TGRS 2012
Relevance in spatial domain Classifier to capture spatial relevance Spatial-Contextual SVM [1] “A Spatial–Contextual Support Vector Machine for Remotely Sensed Image Classification” TGRS 2012
Relevance in spatial domain Classifier to capture spatial relevance Spatial-Contextual SVM
Relevance in spatial domain Classifier to capture spatial relevance Spatial-Contextual SVM
Outline Introduction Relevance in spatial domain Relevance in spectral domain Relevance among multiple features
Relevance in spectral domain Similar spectral properties
Relevance in spectral domain Similar spectral properties
BandClust Splits bands into two disjoint contiguous subbands recursively Splitting criterion: minimizing mutual infromation [2] “BandClust An Unsupervised Band Reduction Method for Hyperspectral Remote Sensing” LGRS 2011 Relevance in spectral domain
BandClust Relevance in spectral domain
CRF to capture spectral domain [3] “ Classification of multitemporal remote sensing data using Conditional Random Fields” PRRS 2010 Relevance in spectral domain
CRF to capture spectral domain [3] “ Classification of multitemporal remote sensing data using Conditional Random Fields” PRRS 2010 Relevance in spectral domain
Outline Introduction Relevance in spatial domain Relevance in spectral domain Relevance among multiple features
Relevance among multiple features Multi-view feature extraction Multi-view classifier One classifier per view, weighted sum of outputs of all classifiers One classifier per view, majority principle Concatenate all features
Relevance among multiple features Multi-view classifier One classifier per view, weighted sum of outputs of all classifiers
Relevance among multiple features Multi-view classifier One classifier per view, weighted sum of outputs of all classifiers
Relevance among multiple features Experiment results
Relevance among multiple features Experiment results PCAGaborlbpsiftConca tenate multiv iew1 multiv iew2 Indian ( ) ( ) ( ) ( ) ( ) ( ) ( ) Flightl ineC ( ) ( ) ( ) ( ) ( ) ( ) ( )