22
Improving Automated Detection of Land Cover Change for Large Areas using High-resolution Satellite Data Kuan Song Department of Geography May 9, 2005

Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

  • Upload
    others

  • View
    2

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Improving Automated Detection of Land Cover Change for Large Areas using High-resolution Satellite Data

Kuan SongDepartment of GeographyMay 9, 2005

Page 2: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Global/Continental/Regional/Local Forest Change

Forest Cover ChangeAn important part of the ‘Global Change’topicLarge amount of forest loss in the last decadeImportant to know how much we lost and how much we still haveFortunate we have quality satellite observations from global to local scale

Page 3: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Global/Continental/Regional/Local Forest Change (illustrations: global)

Global Scale: green indicates forest (Data: 1-km GIMMS)

Forest Loss for logging

Forest Loss for fire

Sept 1990

Sept. 2000

Page 4: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Global/Continental/Regional/Local Forest Change (illustrations: regional)

The country of Paraguay: red indicates forest, Data: 60m/30m Landsat

Page 5: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Global/Continental/Regional/Local Forest Change: Mapping Methods

Old Times: National CensusSpace Age: Satellites

Global/Continental: low/medium-resolution, can tell people where the forest changes occur on this planetRegional/Local: high resolution, can tell people how many acres of forest has been lost

Page 6: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Global/Continental/Regional/Local Forest Change: Mapping Methods

The combination of scale and resolution

Small scale-low resolution: pioneer researches early 1970sLarge scale-low resolution: being done regularly at NASA everyday since 1980ssmall scale- high resolution: since 1990sLarge scale-high resolution: Nothing?

Page 7: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Difficulties of Large scale-high resolution Forest Change Mapping

Data Cost16000 scenes times 600$ per sceneFortunately NASA paid for researchers in 2000

Data Volume16000 scenes times 600MB of data

Processing TimeIn the days when special signal processing machine had to be built for processing images…Even now machine learning algorithms can be computationally expensive

Human InputThis is the hardest partFor high resolution imagery, each has unique radiometry conditions and had to collect training data by handFor each map produced by machine, people have to go over and correct numerous errors by handSounds like Medieval of a Space Age

Page 8: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Possibility of Large scale-high resolution Forest Change Mapping

The first three major difficulties have become past, thanks for technological advances

Only the human input stays a problem

The key is to get good quality map from very limited human input, and only use human input once. No re-visits.

How can we reduce human inputs to the very minimum? Depends on the ability of approaches

Page 9: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Possibility of Machine Learning of Large scale-high resolution Forest Change

Past and Contemporary Machine Learning Methods for small-scale high-resolution forest changes:

Pixel-basedUnsupervised Classification

K-means, ISOCLUS, etc.Supervised Classification

Maximum Likelihood Classification with Gaussian i.i.d. AssumptionDecision Tree, Neural Network, SVM, etc.

Change Vector AnalysisNeighborhood-based

Usually done after pixel-based analysisObject-based

A hybrid of pixel-based analysis, neighborhood=based analysis and polygon operation

Pixel-based methods are the basis of almost all approaches

Page 10: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Possibility of Machine Learning of Large scale-high resolution Forest Change

The literatures listed numerous applications of using classifiers for detecting forest changesBut most literatures don’t mention how much training data was used/was required for generating a certain level of accuracyTherefore we can only test ourselves to see how well those machine learning methods will perform with a tight constraint of training dataThe test field is in Paraguay, where we have plenty of ground data to validateTested methods include ISOCLUS, Maximum Likelihood Classification, Maximum a posteri Classification, Decision Tree, and SVM

Page 11: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Test of classifiers when training data are abundant

When using plenty of training data, every supervised classifier performed very well, ranging from 80-90% accuracyUnsupervised Classifiers needs different training data to classify every pixel

Page 12: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Test of classifiers when training data are very limited

Training data are selected only out of 20km-20km blocks on latitude/longitude intersectionsOnly decision tree and SVM can achieve useable results

Page 13: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Classifiers with different training data budgets

Table IAccuracy of different classifiers under same setting

85%92%Contextual SVM with MRF

85%92%SVM

80%90%DT

60%-70%88%Contextual MLC with MRF

60%-70%86%Bayesian

60%-70%85%MLC

High proportion of pixels remain unclassified without additional training data.

High proportion of pixels remain unclassified without additional training data.

Automated ISODATA

95%95%DMG Recursive ISODATA + Manual Editing

Accuracy for the change class in multiple-scene test

Accuracy for the change class in single-scene test

Method

Page 14: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Decision Tree and SVM

Pruning of Decision TreeWe pruned the tree from 7500 nodes to 1500 nodes to avoid overtuning, but the numbers are chosen quite arbitrarily

Small Training set for SVMThe training set for SVM is very small, in fact 7000 training pixels were used to classify a region with 64M pixels

N-fold cross validation of SVM5-fold CV was used to find best parameter set of the RBF kernel (new slides)

MRF extension of SVMNeighborhood information has been used to derive MAP of SVM (new slides)

Pseudo-data for SVMThe different reasoning between MLC and SVM leads to the criteria of ‘how to guarantee enough training data?’ (new slides)

Page 15: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

N-fold cross validation of SVM

Cross-Validation (CV) is essential in SVM, because an arbitarilychosen parameter of the kernel may give very bad results

Page 16: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

MRF extension of SVM

In the literature there was an MRF extension for MLC, which uses the number of same-class pixels in the neighborhood as prior probabilityFirst experiment in SVM is not really exciting, possibly because the original SVM results are already good.

Ground Truth, SVM Output, Contextual postprocessing of SVM

Page 17: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Pseudo-data for SVM (1)

The assumption of Gaussian i.i.d. for remotely-sensed data is actually quite close to the truth since the data looks like ellipsoidsThe problem is when training data is, in our case, extremely limited, they could not form meaningful ellipsoidsThis is why SVM shows better results

Center of the ellipsoids in the data

Left: Scatterplot of two bands for 1 million pixels; Right: Scatterplot of two bands for same region, but only 1000 pixels

Page 18: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Pseudo-data for SVM (2)

Each training pixel consists records of two time periods, such as A_1, A_2If we mix two training pixels A and B, making C as A_1-B_2, and D as A_2-B_1For MLC it may be a disaster because the Gaussian distributions would be damagedBut for SVM these new data points provide new support vectors

Page 19: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Pseudo-data for SVM (3)

The approach is to divide the training data into 50 clusters by means of K-means clustering, and combines into 2500 combinations. Any combination that does not have a representation in the original training data gets a pseudo training dataA result of such approach is to introduce a ‘Forest Regrowth’ class, which is defined as newly appeared forest. Thus this approach has the potential of discovering small but important classes hard to collect training data directly

Page 20: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Pseudo-data for SVM (4)

Left: SVM with both raw and pseudo training data; Right: Original SVM with raw training dataSmall bright-green area indicates detected forest regrowth

Page 21: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

Acknowledgments

This research was funded under NASA REASoNAgreement #NNG04GC53A and NASA Grant LUCLC #NAG59337.

Part of the research result in this report has been compiled in the paper:Song, K., Townshend, J.R., Kim, S., Davis, P., Improving Automated Detection of Land Cover Change for Large Areas using Landsat Data, IEEE MultiTemp05 Conference, May 16-18, 2005 Mississippi

Page 22: Improving Automated Detection of Land Cover Change forjoseph/Improving Automated... · 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311

References[1] Hansen, M., DeFries, R., Townshend, J., Carroll, M. Dimiceli, C., and Sohlberg, R., "Global Percent Tree Cover at a Spatial Resolution of 500 Meters: First Results of the MODIS Vegetation Continuous Fields Algorithm", Earth Interactions2003, Vol. 7 Issue 1, p1.[2] Olson, D.M., Dinerstein, E., Wikramanayake, E.D., Burgess, N.D., Powell, G.V.N. Underwood, E.C. D'amico, J.A., Itoua, I., Strand, H.E., Morrison, J.C., Loucks, C.J., Allnutt, T.F., Ricketts, T.H., Kura, Y., Lamoreux, J.F., Wettengel, W.W., Hedao, P., Kassem, K.R., “Terrestrial Ecoregions of the World: A new map of life on Earth”, , BioScience, 2001,Vol.51 No.11, 933-38.[3] Quinlan, J. R., “Induction of Decision Trees”., Machine Learning ,1986, 1(1): 81-106, [4] Burges, C. C., A Tutorial on Support Vector Machines for Pattern Recognition. Data Mining and Knowledge Discovery, 1998, 2, 121-167[5] Vapnik, V., The Nature of Statistical Learning Theory. 1995, Springer-Verlag, New York[6] Besag, J., On the Statistical Analysis of Dirty Picture, J. Royal Statistical Society, Ser. B, 48 (1986), pp. 259-302.[7] Settle, J.J., Contextual Classification: Principles and Practice, in Hardy, J.R., Townshend, J.R.G., Settle, J.J., Drake, N.A., Briggs, S.A. (eds), Proceedings of workshop on Contextual Classification of Remotely Sensed Data, July 1987, Department of Geography, University of Reading, U.K. [8] Rudjer Boskovic Institute, Online Tutorial on Decision Trees, http://dms.irb.hr/tutorial/tut_dtrees.php, Feb 2004[9] LIBSVM, A Library for Support Vector Machine, Chih-Chung Chang and Chih-Jen Lin, http://www.csie.ntu.edu.tw/~cjlin/libsvm/, Feb 2004[10] Huang, C., Townshend, J.R.G., “A Stepwise Regression Tree for nonlinear approximation: application to estimating subpixel land cover”, Int. J. Remote Sensing, 2003, Vol.24, No. 1, 75-90[11] Huang, C., Davis, L. S., Townshend, J.R.G., “An Assessment of support vector machines for land cover classification”, Int. J. Remote Sensing, 2002, Vol. 23, No. 4, 725-749[12] Leckie, D.G. Walsworth, N. Dechka, J. Wulder, M. An investigation of two date unsupervised classification in the context of a national program for Landsat based forest change mapping, Geoscience and Remote Sensing Symposium, 2002. IGARSS '02. 2002 IEEE International, 24-28 June 2002, Volume: 3, page(s): 1307 - 1311 vol.3[13] John Rogan, Janet Franklin and Dar A. Roberts, A comparison of methods for monitoring multitemporal vegetation change using Thematic Mapper imagery, Remote Sensing of Environment, Volume 80, Issue 1 , April 2002, Pages 143-156[14] Mary Pax-Lenney, Curtis E. Woodcock, Scott A. Macomber, Sucharita Gopal and Conghe Song, Forest mapping with a generalized classifier and Landsat TM data, , Volume 77, Issue 3 , September 2001, Pages 241-250[15] Liu, D, Gong, P. Kelly, M., A Markov Random Field approach based on Support Vector Machine for change detection, IEEE MultiTemp05 Conference, May 16-18, 2005 Mississippi