27
Managed by UT-Battelle for the Department of Energy Recent Advances in Crop Classification Raju Vatsavai ([email protected]) Computational Sciences and Engineering Division ORNL, Oak Ridge, TN, USA Collaborators: B. Bhaduri, V. Chandola, G. Jun, J. Ghosh, S. Shekhar, T. Burk Remote Sensing – Beyond Images Workshop, Mexico City, Mexico, 14 th December, 2013.

Recent Advances in Crop Classification

Embed Size (px)

DESCRIPTION

Remote sensing –Beyond images Mexico 14-15 December 2013 The workshop was organized by CIMMYT Global Conservation Agriculture Program (GCAP) and funded by the Bill & Melinda Gates Foundation (BMGF), the Mexican Secretariat of Agriculture, Livestock, Rural Development, Fisheries and Food (SAGARPA), the International Maize and Wheat Improvement Center (CIMMYT), CGIAR Research Program on Maize, the Cereal System Initiative for South Asia (CSISA) and the Sustainable Modernization of the Traditional Agriculture (MasAgro)

Citation preview

Page 1: Recent Advances in Crop Classification

Managed by UT-Battellefor the Department of Energy

Recent Advances in Crop Classification

Raju Vatsavai([email protected])

Computational Sciences and Engineering Division

ORNL, Oak Ridge, TN, USA

Collaborators:

B. Bhaduri, V. Chandola, G. Jun, J. Ghosh, S. Shekhar, T. Burk

Remote Sensing – Beyond Images Workshop, Mexico City, Mexico,

14th December, 2013.

Page 2: Recent Advances in Crop Classification

2 Managed by UT-Battellefor the Department of Energy

Outline

· Better spectral and spatial resolution– Fine-grained (species) classification– Complex (compound) object recognition

· Challenges– Limited ground-truth: Semi-supervised learning (SSL)– Spatial homogeneity: SSL + Markov Random Fields– Spatial heterogeneity: Gaussian Process (GP) learning– Aggregate vs. Subclasses: Fine-grained classification– Phenology: Multi-view learning

· Conclusions

Page 3: Recent Advances in Crop Classification

3 Managed by UT-Battellefor the Department of Energy

Challenge 1: Limited Training Data

· Increasing spectral resolution: 4 to 224 Bands

· Challenges– #of training samples ~ (10 to 30) * (number of dimensions)– Costly ~ $500-$800 per plot (depends on geographic

area)– Accessibility – Private/Privacy issues (e.g., USFS may

average 5% denied access)– Real-time – Emergency situations, such as, forest fires,

floods

· Solutions– Reduce number of dimensions– (Artificially) Increase number of samples– By incorporating unlabeled samples

· Naïve semi-supervised (Nigam et al. [JML-2000])– Bagging [Breiman, ML-96]

Page 4: Recent Advances in Crop Classification

4 Managed by UT-Battellefor the Department of Energy

True Distribution

Estimated Distribution(Small Samples; MLE are good asymptotically)

Page 5: Recent Advances in Crop Classification

5 Managed by UT-Battellefor the Department of Energy

Initial Estimates +Unlabeled Samples

Page 6: Recent Advances in Crop Classification

6 Managed by UT-Battellefor the Department of Energy

Iteratively Update Parameters Using Unlabeled Samples

Page 7: Recent Advances in Crop Classification

7 Managed by UT-Battellefor the Department of Energy

Iteratively Update Parameters Using Unlabeled Samples

Page 8: Recent Advances in Crop Classification

8 Managed by UT-Battellefor the Department of Energy

Iteratively Update Parameters Using Unlabeled Samples

Page 9: Recent Advances in Crop Classification

9 Managed by UT-Battellefor the Department of Energy

Final parameters after convergence

Page 10: Recent Advances in Crop Classification

10 Managed by UT-Battellefor the Department of Energy

· E-Step

· M-Step

ithdata vector, jth class

Solution: Semi-supervised Learning

Assume Samples are generated by a Gaussian Mixture Model (GMM)

• Estimate Parameters with Expectation Maximization (EM)

Page 11: Recent Advances in Crop Classification

11 Managed by UT-Battellefor the Department of Energy

Results

10 Classes, 100 Training Samples(10-30) x No of dimensions / class

Small Subset of 20 Training Samples

20 labeled + 80 unlabeled samples

Supervised (BC) vs. Semi-supervised (BC-EM)

Fixed Unlabeled (85) and Varying (Increasing) Labeled

0 20 40 60 80 100 120

Acc

urac

y

30

40

50

60

70

80

BC - WorstBC - BestBC (EM) - Best

Ranga Raju Vatsavai, Shashi Shekhar, Thomas E. Burk: A Semi-Supervised Learning Method for Remote Sensing Data Mining. ICTAI 2005: 207-211

Page 12: Recent Advances in Crop Classification

12 Managed by UT-Battellefor the Department of Energy

Prior Distribution Model:· For Markov random field , the conditional

distribution of a point in the field given all other points is only dependent on its neighbors.

s excluding Sin points ofset a denotes

lattice imagean is Where

)}(|)({)}(|)({

sS

S

sspsSsp

sxx

xx

sxx

xxx x

x xsxx

xxx

x x

xx

x

xx

Challenge 2: Spatial Homogeneity

Bayes Theorem: p(c|x) = p(x|c)p(c)/p(x)

For a first - order neighborhood system

p() 1

ze

t c ( )C

e.q.1

t c () is the total number of horizantally

and vertially neighboring points of different

value in in clique c.

e.q.1 is Gibbs distribution and therefore,

an MRF.

is emphirically determined weight.

t c () {0, otherwise. 1 if (i, j ) (k,l )

Spatial Homogeneity

Page 13: Recent Advances in Crop Classification

13 Managed by UT-Battellefor the Department of Energy

Solution: Spatial Classification

BC (60%) BC-EM (68%)

BC-MRF (65%) BC-EM-MRF (72%)

• Shashi Shekhar, Paul R. Schrater, Ranga Raju Vatsavai, Weili Wu, Sanjay Chawla: Spatial contextual classification and prediction models for mining geospatial data. IEEE Transactions on Multimedia 4(2): 174-188 (2002)

• Baris M. Kazar, Shashi Shekhar, David J. Lilja, Ranga Raju Vatsavai, R. Kelley Pace: Comparing Exact and Approximate Spatial Auto-regression Model Solutions for Spatial Data Analysis. GIScience 2004: 140-161

• Ranga Raju Vatsavai, Shashi Shekhar, Thomas E. Burk: An efficient spatial semi-supervised learning algorithm. IJPEDS 22(6): 427-437 (2007)

Page 14: Recent Advances in Crop Classification

14 Managed by UT-Battellefor the Department of Energy

Challenge 3: Spatial Heterogeneity· Going From Local to Global

– Signature continuity is a problem in classifying large geographic regions

· Solutions– Assume constant variance structure over space, that is,

train one model, use it on other regions – poor performance

– Train separate model for each region – needs lot of data– Train one model covering samples from all regions –

needs an adaptive model to capture spatial heterogeneity

Page 15: Recent Advances in Crop Classification

15 Managed by UT-Battellefor the Department of Energy

Solution: Gaussian Process (GP) Classification

· Change of distribution over space is modeled by),(~)|( Nyxp

))(),((~)|)(( ssNysxp

Goo Jun, Ranga Raju Vatsavai, Joydeep Ghosh: Spatially Adaptive Classification and Active Learning of Multispectral Data with Gaussian Processes. SSTDM 2009: 597-603

Page 16: Recent Advances in Crop Classification

16 Managed by UT-Battellefor the Department of Energy

Challenge 4: Aggregate Vs. Sub-classes

· Spectral Classes vs. Thematic Classes

· Insufficient Ground-truth· Subjective/domain-dependent· Parametric – assumption violations

Page 17: Recent Advances in Crop Classification

17 Managed by UT-Battellefor the Department of Energy

Solution: Sub-class Classification

· Coarse-to-fine Resolution Information Extraction– Characterizing the nature of the change

· Fallow to Switch grass, Wheat to Corn, or crop damage

Coarse Classes (MODIS)Each class is Gaussian

Sub-Classes (AWiFS)Each class is MoG

Model Selection (BIC,AIC)How many components?Parameter Estimation

Semi-supervised Learning

Characterize Changes

Page 18: Recent Advances in Crop Classification

18 Managed by UT-Battellefor the Department of Energy

Results: Sub-class Classification

Dataset:LandSat ETM+ Data (Cloquet, Carleton, MN, May 31, 2000)• 6 Bands, 4 Classes, 60 plots• Independent test data: 205 plots• Forest (4 Subclasses; 2 subclasses are

combined into 1)• 2 Labeled plots per sub-class

1. Ranga Raju Vatsavai, Shashi Shekhar, Budhendra L. Bhaduri: A Learning Scheme for Recognizing Sub-classes from Model Trained on Aggregate Classes. SSPR/SPR 2008: 967-976

2. Ranga Raju Vatsavai, Shashi Shekhar, Budhendra L. Bhaduri: A Semi-supervised Learning Algorithm for Recognizing Sub-classes. SSTDM 2008: 458-467

Page 19: Recent Advances in Crop Classification

19 Managed by UT-Battellefor the Department of Energy

Crop (Opium) Classification

· Helmand accounts for 75% of the world’s opium production

· GeoEye 4-Band Image, 13th May 2011

Page 20: Recent Advances in Crop Classification

20 Managed by UT-Battellefor the Department of Energy

Ground-truth (Aggregate Classes)

· Ground-truth collected for 4 classes

· 1-Other Crops (Yellow), 2-Poppy (Red), 3-Soils (Cyan), 4-Water (Blue)

Page 21: Recent Advances in Crop Classification

21 Managed by UT-Battellefor the Department of Energy

Classified (Aggregate) Image

· Maximum Likelihood Classification (Widely used)

· Also did lot of other standard classification schemes – Decision Trees, Random Forest, Neural Nets, …

Page 22: Recent Advances in Crop Classification

22 Managed by UT-Battellefor the Department of Energy

Classified (Sub-classes) Image

· Sub-class classification – Identifying finer classes from aggregate class – new scheme– 1 -> 11,12,13; 2 -> 21,22,23, 3->31,32, 4->41

· (Overall Accuracy Improved by ~10%)

Page 23: Recent Advances in Crop Classification

23 Managed by UT-Battellefor the Department of Energy

Challenge 5: Phenology

AWiFS (May 3, 2008; FCC (4,3,2))

AWiFS (July 14, 2008; FCC (4,3,2))

Thematic Classes: C-Corn, S-Soy

Page 24: Recent Advances in Crop Classification

24 Managed by UT-Battellefor the Department of Energy

More Formally

Page 25: Recent Advances in Crop Classification

25 Managed by UT-Battellefor the Department of Energy

Solution: Multi-view Learning

· Multi-temporal images are different views of same phenomena– Learn single classifier on different views, chose

the best one through empirical evaluation– Combine different views into a single view, train

classifier on single combined view – stacked vector approach

– Learn classifier on single view and combine predictions of individual classifiers – multiple classifier systems

· Bayesian Model Averaging

– Co-training· Learn a classifier independently on each view· Use predictions of each classifier on unlabeled

data instances to augment training dataset for other classifier

Varun Chandola, Ranga Raju Vatsavai: Multi-temporal remote sensing image classification - A multi-view approach. CIDU 2010: 258-270

Page 26: Recent Advances in Crop Classification

26 Managed by UT-Battellefor the Department of Energy

Conclusions· We developed several innovative solutions that

address big spatiotemporal data challenges– Semi-supervised learning– Spatial classification (homogeneity and heterogeneity)– Temporal classification– Sub-class classification

· Ongoing– Transfer learning: Adopt model learned in area to the

other with very little additional ground-truth– Compound object classification (multiple instance

learning)– Semantic classification (beyond pixels and objects)– Scaling

· Heterogeneous (OpenMP + MPI + CUDA)· Cloud computing (MapReduce)

Page 27: Recent Advances in Crop Classification

27 Managed by UT-Battellefor the Department of Energy

Acknowledgements

· Prepared by Oak Ridge National Laboratory, P.O. Box 2008, Oak Ridge, Tennessee 37831-6285, managed by UT-Battelle, LLC for the U. S. Department of Energy under contract no. DEAC05-00OR22725.

· Collaborators and Sponsors