Arnt-Børre Salberg and Rune Solberg Norwegian Computing Center

www.nr.noearthobs.nr.no

Land cover classification of cloud- and snow-contaminated multi-temporal high-resolutionsatellite images

Arnt-Børre Salberg and Rune SolbergNorwegian Computing Center

3rd Workshop of the EARSeL SiG Remote Sensing of Land Use and Land Cover, 25 - 27 November 2009, Bonn, Germany


Overview

► Motivation & challenges

► Missing data mechanism

► Classification with missing observations

► Image restoration

► Experiments & Results

► Summary and Discussions


Motivation –Land cover classification

Classifier

Multi-spectral image Thematic map

Featurevector

Label


Multi-temporal land cover classification► Land cover classification using high-resolution

optical remote sensing can be challenging since:▪ In Northern Europe clouds and snow prevent us

from observing the surface of the earth.▪ High-resolution images has often a low temporal

coverage.

► Multi-temporal land cover classification▪ Enhanced performance since we observe the

vegetation at different phenological states.▪ The set of cloud contaminated images have

observed a higher portion of the earth’s surface than a single image.


Multi-temporal land cover classification by pixel level fusion

Multi-temporal & Multi-spectral imagesThematic map

Featurevector Label

Classifier


Challenges – Pixel level fusion?

Image 1 X X X X

Image 2 X X X X X

Image 3 X X X X X X

Pixel no. 1 2 3 4 5 6 7 8 9 10 11 12

Typical missing data pattern

► How should we handle the missing observations?


Handling missing observations

Proposed approach:► Identify the missing observations.

► Identify the missing data mechanism.

► Construct classifiers capable of handling data with missing features and a given missing data mechanism.


Identify missing observations

► Cloud/snow detection▪ Classify the images into the categories: Cloud,

snow, water and vegetation/soil/rock.▪ Constructed a missing data indicator ri for each

pixel

► Assume perfect cloud/snow detection


Identify the missing data mechanisms► Missing completely at random (MCAR)

▪ Landsat 7 sensor failure.

► Missing at random (MAR)▪ Clouds

► Not MAR▪ Snow, censoring of measurements


Classification with missing observations

Some existing approaches► Mean value or zero substitution

▪ Biased estimates

► Remote sensing▪ Aksoy et al. 2009 ▪ Decision tree based approach


Classification with missing observationsLet x(k) denote the part of x corresponding to the missing data indicator vector rk

Optimal classifier (Mojirsheibani & Montazeri, 2007)

Let be a binary vector with 0 at the element j if the jth element of x is missing, and 1 otherwise


Classification with missing observations

► Missing data mechanism introduces an additional probability

▪ Depends on feature vector and land cover class.

► MCAR:

▪ Classifier reduces to the marginal distribution where the missing features are integrated out.


► Unknown parameters need be estimated when applying parametric classifiers

► Only use complete feature vectors for learning ▪ May be only a few available

► Expectation Maximization algorithm often applied for Gaussian distributions or mixture Gaussian distributions

► Parametric classifiers difficult since

is unknown and hard to estimate.

Parametric classifiers


► K-NN classifier for not MAR scenarios:

▪ kNN classifier works on the selection of samples among the training data that has the exact same missing data pattern as the test vector, and perform the kNN rule among these samples

Non-parametric classifiers


Two-stage classifier


► Assume that a land cover map is available (from the classification module)

► Minimum mean-squared error estimator (assuming Gaussian distributions)

▪ Dependent on the land cover class of the given pixel. c and c estimated using the EM algorithm (MAR

assumption)

Image restoration


Experiments & Results

► Land cover classification of mountain vegetation important for biomass estimation of lichen.▪ Remote sensing data: 4 Landsat 7 ETM+ images

(2004-05-31, 2000-07-23, 2002-08-14, and 2002-09-15)

▪ Ancillary data: Slope and elevation derived from a digital elevation model (DEM).

▪ In situ data: 4861 pixels were labeled according to the classes: water, ridge, leeside, snowbed, mire, forest and rock.


Results– Land cover classification

Input images Missing data indicators Thematic map


Results – Cloud removal

Input image Restored imageCloud shadows

► Image restoration of July 23 using Aug. 14 and Sep. 15 images.


Results – snow and sensor failure removal

Input image Restored image

► Image restoration of May 31 image using July 23, Aug. 14 and Sep. 15 images.

► Note that at May 31 the vegetation is in a different phenological state than for the other images.


Classification resultsMethod July 23

2000Aug 14 2002

Sep. 15 2002

DEM Acc. excl. missing data

Acc. incl. missing data

Portion classified

Gauss. X 69% 52% 63%

X 63% 57% 76%

X 67% 66% 99%

X X X X 78% 78% 100%

K-NN X 68% 51% 63%

X 63% 57% 76%

X 67% 66% 99%

X X X X 81% 81% 100%


Summary and discussions

► Proposed a two-stage approach▪ Cloud/snow classification▪ Vegetation type classification with missing observations

► Obtained increased classification power by pixel level fusion of cloud and snow contaminated satellite images

► Image restoration natural by product and seem to work good for some areas.▪ Cloud shadows remains a challenge.▪ Difficult for not MAR


Summary and discussions

► Further improvement in classification accuracy expected by▪ Proper feature extraction▪ Contextual classification (e.g. Markov Random Field)▪ Including ancillary data important for mountain

vegetation (e.g. bio-climatic variables)▪ Multi-sensor fusion with full polarimetric SAR

images?▪ Identification of cloud shadows▪ Topographic illumination correction (c-correction)


Thank you

Arnt-Børre Salberg and Rune Solberg Norwegian Computing Center

Documents