Presentation is loading. Please wait.

Presentation is loading. Please wait.

IMPROVING ACTIVE LEARNING METHODS USING SPATIAL INFORMATION IGARSS 2011 Edoardo Pasolli Univ. of Trento, Italy Farid Melgani Univ.

Similar presentations


Presentation on theme: "IMPROVING ACTIVE LEARNING METHODS USING SPATIAL INFORMATION IGARSS 2011 Edoardo Pasolli Univ. of Trento, Italy Farid Melgani Univ."— Presentation transcript:

1 IMPROVING ACTIVE LEARNING METHODS USING SPATIAL INFORMATION IGARSS 2011 Edoardo Pasolli Univ. of Trento, Italy pasolli@disi.unitn.it Farid Melgani Univ. of Trento, Italy July 28, 2011 Devis Tuia Univ. of València, Spain Fabio Pacifici DigitalGlobe, Colorado William J. Emery Univ. of Colorado at Boulder

2 Introduction  Supervised classification approach 2 Pre- processing Feature extraction Classification Image/ Signal Decision Training sample collection Training sample quality/quantity Human expert Impact on accuracies

3 Introduction  Active learning approach 3 Training of classifier Active learning method Model of classifier Learning (unlabeled) set Labeling of selected samples Selected samples after labeling Insertion in training set f1f1 f2f2 f1f1 f2f2 f1f1 f2f2 Selected samples from learning (unlabeled) set f2f2 f1f1 f2f2 f1f1 Human expert Training (labeled) set Class 1 Class 3 Class 2

4 Objective  Propose SVM-based active learning strategy for classification of remote sensing images by combining spectral and spatial information 4

5 Support Vector Machines (SVMs)  Training set:  Kernel function:  Dual optimization problem maximize subject to 5

6 Support Vector Machines (SVMs)  Discriminant function 6 SVM Training f1f1 f2f2 Training (labeled) set in feature space Class 1 Class 2 f1f1 f2f2 Training (labeled) set in feature space SVM model 0absolute value of discriminant function : SV

7 Proposed Strategy 7 L: Training set SV s : Support vectors SVM Training Spectral selection criterion Spatial selection criterion U ’ s : Selected samples Selection Insertion in training set U s : Sorted samples U: Learning set Human expert Nondominated sorting Labeling L ’ s : Labeled samples

8 Spectral Criterion: Margin Sampling (MS) 8 Selection Learning (unlabeled) set in feature space f1f1 f2f2 f1f1 Training (labeled) set in feature space f2f2 SVM model f1f1 f2f2 Selected samples from learning (unlabeled) set in feature space selection of samples with minimum absolute values of discriminant function 0absolute value of discriminant function : SV

9 Spatial Criterion: Distance from SVs (Sp) 9 Selection selection of samples with maximum distance values from the closest SV : SV 0- distance value from the closest SV f1f1 Training (labeled) set in feature space f2f2 SVM model Learning (unlabeled) set in spatial space Selected samples from learning (unlabeled) set in spatial space Training (labeled) set in spatial space

10 Combined Criterion (MS&Sp) 10 determined by nondominated sorting selection of samples starting from the Pareto Front 1front number Front 1: Pareto Front Front 2 Front 3 Front 4 Front 5 MS: absolute value of discriminant function Sp: - distance value from the closest SV

11 Experimental Results  Data set description  Test site: Las Vegas, Nevada  Acquisition date: 2002  Sensor: QuickBird  # features: 4 spectral + 36 morphological  Spatial resolution: 0.6 m  # thematic classes: 11 11 False color compositing Commercial buildings Residential houses Drainage channel Roads Trees Short vegetation Water Bare soil Parking lots Soil Highways

12 Experimental Results  Data set description  Test site: Las Vegas, Nevada  Acquisition date: 2002  Sensor: QuickBird  # features: 4 spectral + 36 morphological  Spatial resolution: 0.6 m  # thematic classes: 11 12 Ground truth Commercial buildings Residential houses Drainage channel Roads Trees Short vegetation Water Bare soil Parking lots Soil Highways

13 Experimental Results 13 Training setMS criterion Sp criterionMS+Sp criterion 0 absolute value of discriminant function 0 - distance value from the closest SV 1 front number

14 Experimental Results  Overall Accuracy and Kappa index 14

15 Experimental Results  Absolute value of discriminant function normalization of map standard deviation 15

16 Experimental Results  Detailed results 16 Accuracies on 343,023 test samples Method # training samples OA σ OA Kappa σ Kappa AA σ AA σ DF Full3000095.47-0.947-93.35-0.06 Initial5558.985.740.5330.0659.334.100.45 R MS MS&Sp 1035 84.89 88.09 89.73 0.56 0.40 0.24 0.823 0.860 0.880 0.007 0.005 0.003 79.22 83.15 84.86 1.47 0.80 0.38 0.25 0.16 R MS MS&Sp 2035 87.18 90.54 92.13 0.42 0.89 0.19 0.850 0.889 0.908 0.005 0.01 82.00 86.55 88.39 1.00 1.07 0.27 0.22 0.18 0.16

17 Experimental Results  Detailed results 17 Accuracies on 343,023 test samples Method # training samples OA σ OA Kappa σ Kappa AA σ AA σ DF Full3000095.47-0.947-93.35-0.06 Initial5558.985.740.5330.0659.334.100.45 R MS MS&Sp 1035 84.89 88.09 89.73 0.56 0.40 0.24 0.823 0.860 0.880 0.007 0.005 0.003 79.22 83.15 84.86 1.47 0.80 0.38 0.25 0.16 R MS MS&Sp 2035 87.18 90.54 92.13 0.42 0.89 0.19 0.850 0.889 0.908 0.005 0.01 82.00 86.55 88.39 1.00 1.07 0.27 0.22 0.18 0.16

18 Experimental Results  Detailed results 18 Accuracies on 343,023 test samples Method # training samples OA σ OA Kappa σ Kappa AA σ AA σ DF Full3000095.47-0.947-93.35-0.06 Initial5558.985.740.5330.0659.334.100.45 R MS MS&Sp 1035 84.89 88.09 89.73 0.56 0.40 0.24 0.823 0.860 0.880 0.007 0.005 0.003 79.22 83.15 84.86 1.47 0.80 0.38 0.25 0.16 R MS MS&Sp 2035 87.18 90.54 92.13 0.42 0.89 0.19 0.850 0.889 0.908 0.005 0.01 82.00 86.55 88.39 1.00 1.07 0.27 0.22 0.18 0.16

19 Conclusions  In this work, new SVM-based active learning strategy by combining spectral and spatial information is proposed  Encouraging performances in terms of  classification accuracy: convergence speed and stability  classification reliability  Drawbacks  higher computational load 19

20 IMPROVING ACTIVE LEARNING METHODS USING SPATIAL INFORMATION IGARSS 2011 Edoardo Pasolli Univ. of Trento, Italy pasolli@disi.unitn.it Farid Melgani Univ. of Trento, Italy July 28, 2011 Devis Tuia Univ. of València, Spain Fabio Pacifici DigitalGlobe, Colorado William J. Emery Univ. of Colorado at Boulder


Download ppt "IMPROVING ACTIVE LEARNING METHODS USING SPATIAL INFORMATION IGARSS 2011 Edoardo Pasolli Univ. of Trento, Italy Farid Melgani Univ."

Similar presentations


Ads by Google