Statistical alignment of remote sensing images to improve classifiers portability and change detection

Transcription

1 Statistical alignment of remote sensing images to improve classifiers portability and change detection Institute of Geomatics and Analysis of Risk University of Lausanne, Switzerland 7th of June Journée LSSR Yverdon

2 Motivation Image classification: Impossible to carry out terrain campaigns to obtain ground truth for every acquired image. Intelligently reuse the information provided by labeled pixels from similar images. Change detection: Ambiguity problem between changed and unchanged pixels. Better representation of the images to highlight changed regions.

3 Motivation Issues when analyzing multiple remote sensing images: different illumination, changing atmospheric conditions, varying acquisition geometry, seasonal effects,... shifted probability distributions between images. need to match/align the images: physical models (atmospheric correction), histogram matching,...

4 A concrete example: QuickBird images of Zurich Source image (autumn) Target image (summer)

5 Ground truth Source image (autumn) Target image (summer)

6 Red vs NIR scatterplots Source training set Target test set.5 buildings.5 buildings grass grass vegetation vegetation roads roads shadows shadows.5.5 standardized NIR band.5 standardized NIR band standardized R band standardized R band

7 Objectives Domain Adaptation via feature extraction Map the two images into a feature space where the differences are reduced. Test different feature extraction methods (PCA, KPCA, TCA). Image classification: Apply on the target image a classifier built on labeled source samples only. Change detection: Enhanced quality of the difference image.

8 Domain Adaptation via feature extraction: principle D S = {X S, Y S } labeled source training data. X T unlabeled target data. Find a common mapping φ (feature extraction techniques: PCA, KPCA, etc.): X S φ(x S ) = X S, X T φ(x T ) = X T. Reduce differences between distributions so that P(X S ) P(X T ). original source image X S feature extraction based on X = X S X T mapping φ embedded source image X* S prediction training on X* T using source labels Y S supervised classifier original target image X T mapping φ embedded target image X* T final target ground cover map Y T

9 Classification accuracy on target image κ Target image LDAtgt: K =.79 LDAsrc: K =.54 LDA_TCA: K = LDAtgt LDAsrc LDA_PCA LDA_TCA Nr. of features Gain of.5 κ points over the Source model (average over runs).

10 Extracted features.5.5 buildings grass vegetation roads shadows.5.5 buildings grass vegetation roads shadows Original R vs NIR standardized NIR band.5.5 standardized NIR band standardized R band standardized R band Transformed TCA comps. 3 vs 4 Comp Comp. 3 Source image Comp Comp. 3 Target image

11 A multitemporal application: change detection Definitions Pixel-based comparison of (co-)registered images to detect spectral differences related to ground cover changes. Image differencing (Change Vector Analysis) is the most applied technique: Magnitude of difference pixel vector < θ No change. Magnitude of difference pixel vector > θ Change.

12 Introduction Domain Adaptation Image classification Change detection Summary Image alignment in change detection Aligning unchanged areas Based on some no change information (easy to obtain) match the distribution of images pre- and post-event. A common set of unchanged pixels, at same locations in both images, is used to extract the new projection. Physical meaning is lost, but pixel-wise comparison is improved. t t GT t transf. t transf.

13 Introduction Domain Adaptation Image classification Change detection Summary 3 NIR 3 3 R 3 original magnitude 4 5 kpca difference of component Enhanced difference image representation kpca difference of component 6 transformed magnitude Even if the magnitude looks noisy the separability between changes / no changes is increased!

14 Change detection results original transformed CVA RBF SVDD Est. κ statistic Retained Principal Components transformed CVA transformed RBF SVDD original CVA original RBF SVDD Improvements over the original space with different classification techniques:. κ points (average over runs).

15 Summing up Feature extraction techniques efficiently align images in the feature space. Image classification: newly acquired images can be suitably classified using already existing ground truth. classifiers portability Change detection: the projection aligns unchanged pixels emphasizing changed regions. enhanced changed detection

16 The end Thank you for your attention! Any questions? This work has been supported by the Swiss National Science Foundation with grants no and PZP-3687.