Textural analysis of coca plantations using 1-meter-resolution remotely-sensed data

Transcription

1 UNODC Workshop, November, Bogota, Colombia 1 Textural analysis of coca plantations using 1-meter-resolution remotely-sensed data Workshop on Measurement of Cultivation and Production of Coca Leaves and Derivates United Nations Office on Drugs and Crime (UNODC) November 2008, Bogota, Colombia Martino Pesaresi European Commission, Joint Research Centre, Institute for the Protection and Security of the Citizen (IPSC) Support to External Security Unit (SES) Via E. Fermi, 1 TP 267 Ispra (VA), Italy Tel (direct line) Fax martino.pesaresi@jrc.it

2 Index UNODC Workshop, November, Bogota, Colombia 2 ISFEREA Satellite Image Intelligence Very High Resolution Satellite Data Automatic Image Understanding Image Information Mining Geographic Information Systems Spatial Decision Support Systems Collaborative Mapping Systems Spatial Data Repositories Human Security Settlement Analysis Informal settlements Refugee and IDP Camps Population Risk Assessment Damage and Reconstruction Assessment Conflict Modeling Conflict Resources Monitoring Extraction of information from VHR data Image textural analysis The concept of mature coca index (MCI) Application of the MCI concept for monitoring coca crops

3 Textural analysis steps UNODC Workshop, November, Bogota, Colombia 3 Data input: ikonos panchromatic 1m spatial resolution Reference information Anisotropic analysis Separability analysis Multi-scale accuracy analysis Dimensional reduction Mature Coca Index class hectares percent Mature coca % Forest % Clearing % Shadow % total %

4 UNODC Workshop, November, Bogota, Colombia 4 A 380x530 meters subsample of the data input used for this experiment. A) mature coca field, B) forest, C) clearing areas, and D) shadows.

5 Radiometric Information UNODC Workshop, November, Bogota, Colombia % 1.80% 1.60% 1.40% 1.20% 1.00% 0.80% 0.60% coca forest clear shadow 0.40% 0.20% 0.00% DN

6 What is texture (GLCM) UNODC Workshop, November, Bogota, Colombia 6 r tx = f d, DIS = ASM = ( w v m) Ng,,α Ng i= 1 j= 1 Ng Ng P i, j i j ( P ) i, j i= 1 j= 1 Ng Ng i= 1 j= 1 2 ( i j) 2 Pi j CON, = Ng Ng IDM = P 2 i, j i= 1 j= ( i j) Ng Ng ENT = P i, j log Pi, j i= 1 j= 1 ( ) Grey Level Co-occurrence Matrix (GLCM) contains the relative frequencies with which two neighboring pixels (linked by a spatial relation) occur on the local domain (window) of the image, one with gray level I and the other with gray level J. angle = 270 distance angle = 180 x angle = 0,360 angle = 90

7 2500 Anisotropic Analysis UNODC Workshop, November, Bogota, Colombia 7 B4Stdv B5Stdv r tx = f d, ( w v m),,α Stdv d=1 d=2 d=3 d=4 d=5 Stdv d=1 d=2 d=3 d=4 d= Angle Angle Figure 1 The anisotropic analysis of the contrast (left) and dissimilarity (right) co-occurrence textural measures. B3Stdv B6Stdv B7Stdv Stdv d=1 d=2 d=3 d=4 d=5 Stdv d=1 d=2 d=3 d=4 d=5 Stdv d=1 d=2 d=3 d=4 d= Angle Angle Angle Figure 2 From left to right the anisotropic analysis of the homogeneity, entropy and second moment co-occurrence textural measures

8 Separability Analysis UNODC Workshop, November, Bogota, Colombia 8 Separability Analysis r tx = f d, ( w v m),,α J e ffrie s -M a tu s ita homogeneity contrast dissimilarity entropy second moment c oca&forest forest&clear coca&clear shadow &clear Class Pairs s hadow &fores t shadow&coca Av erage Figure 1 Separability analysis of the selected co-occurrence textural measures

9 W size and accuracy r tx = f d, ( w v m),,α UNODC Workshop, November, Bogota, Colombia 9 Producer Accuracy User Accuracy % Overall Accuracy coca forest shadow clear % Overall Accuracy coca forest shadow clear Window Size Window Size Commission Error Omission Error % Overall Accuracy coca forest shadow clear % Overall Accuracy coca forest shadow clear Window Size Window Size Figure 1 - Results of the accuracy analysis using contrast textural measure, α parameter d distance parameter. equal to 45 degrees, and { }

10 Scale (distance) and Classes, MCI definition UNODC Workshop, November, Bogota, Colombia 10 Contrast, 45deg 4000 r tx = f d, ( w v m),,α Mean, Stdv coca forest clear shadow MCI w = ( w tx ) tx d = 1 w d = Distance

11 Contrast, d=1, w=51 UNODC Workshop, November, Bogota, Colombia 11 Textural feature calculated in the same area. Linear min-max stretching has been applied for visualization.

12 Contrast, d=5, w=51 UNODC Workshop, November, Bogota, Colombia 12 Textural feature calculated in the same area of Figure 11. Linear minmax stretching has been applied for visualization.

13 MCI UNODC Workshop, November, Bogota, Colombia 13 MCI calculated using and in the same area of Figure 11. Linear minmax stretching has been applied for visualization.

14 Application pilot study UNODC Workshop, November, Bogota, Colombia 14 It involves 3 test sites of 10x10 kilometers each Test sites are in the area of Magdalena Medio, Cantagallo, San Pablo, Simiti-Santa Rosa It is connected to the activity of ECfounded project Programa Desarrollo y Paz del Magdalena Medio

15 Coca cultivations UNODC Workshop, November, Bogota, Colombia Young coca plants in recently-burned background 2. Mature coca field from the field survey 3. Mature coca fields as seen from satellite image processed by JRC IPSC 1 2 3

16 Data Sources UNODC Workshop, November, Bogota, Colombia 16 Satellite Imagery Landsat TM/ETM, Ikonos Already existing maps INSTITUTO GEOGRÁFICO AGUSTÍN CODAZZI-IGAC 1: Digital Terrain Model NASA SRTM 90m spatial resolution Field survey GIS/GPS portable devices Reference and ancillary data UNODC Sistema Integrado de Monitoreo de Cultivo Ilícitos-SIMCI USAID Pan-American Development Foundation-FUPAD

17 Satellite image processing characteristics UNODC Workshop, November, Bogota, Colombia 17 Multiple spatial/spectral resolution IKONOS multispectral 4 meters IKONS panchromatic 1 meter Landsat TM and ETM multispectral 30 meters Landsat ETM panchromatic 15 meters Multiple date IKONOS 3 dates, 6 months tot Landsat 4 dates, 4 years tot Multiple criteria Spectral signature Textural signature Contextual information Time persistence / change patterns

18 Automatic Image Understanding UNODC Workshop, November, Bogota, Colombia 18 Image fusion for improving visual exploration and interpretation Emphasis on robust automatic detection Multiple radiometric and textural signatures to fuzzy multicriteria Multi-source spatial integration using object-based image processing approach Intensive use of contextual information and a priori knowledge General image understanding using rule-based fuzzy inferential model

19 Data Flow UNODC Workshop, November, Bogota, Colombia 19 SRTM Spatial improvement Elevation Contours generation Ikonos MS 3 dates Ortho rectification Fusion Maps, reports Ikonos PAN 3 dates Mature Coca texture index Ortho rectification Coca Texture 3 dates Fuzzy Inferential Model Image Interpretation digitalization Roads, buildings Field survey, Image interpretation Ground Truth Validation

20 Spectral info of a coca field UNODC Workshop, November, Bogota, Colombia 20

21 Textural info of coca field (from PAN) UNODC Workshop, November, Bogota, Colombia 21 Input Ikonos P Channel Texture CON Co-occurrence Displ. 1,1 Texture CON Co-occurrence Displ. 3,3

22 Mature Coca Index from textural analysis UNODC Workshop, November, Bogota, Colombia 22 Texture CON Co-occurrence Displ. 1,1 Texture CON Co-occurrence Displ. 3,3 Derived Mature Coca Index (d1/d3)

23 Radiometric info and Mature Coca Index UNODC Workshop, November, Bogota, Colombia 23

24 Spatial degradation after orthorectification UNODC Workshop, November, Bogota, Colombia 24

25 Mature Coca Index robust in multi-temporal processing UNODC Workshop, November, Bogota, Colombia 25 MCI Time0 (January) MCI Time1 (March) MCI Time2 (May)

26 Coca detection (in yellow) by region-based image processing UNODC Workshop, November, Bogota, Colombia 26

27 Inferential Model UNODC Workshop, November, Bogota, Colombia 27 Inferential Model Ground Truth Primary input Derived input Ikonos MS t0 Ikonos PAN t0 Mature Coca Index t0 Vegetation Index t0 Static Fuzzy Inferential Model t0 Validation t0 Ikonos MS t1 Ikonos PAN t1 Mature Coca Index t1 Vegetation Index t1 Static Fuzzy Inferential Model t1 Validation t1 Multi-temporal Fuzzy Inferential Model OUT Ikonos MS t2 Ikonos PAN t2 Mature Coca Index t2 Vegetation Index t2 Static Fuzzy Inferential Model t2 Validation t2 Image Segmentation Roads, buildings

28 Static Inferential Model t0,t1,t2 UNODC Workshop, November, Bogota, Colombia 28 6 classes based on radiometric and textural signatures derived from ground truth Mature Coca Young Coca Forest Bare soil No data (clouds or shadows) Other surfaces Validation

29 Multi-temporal Inferential Model 18 classes based on static fuzzy model at t0, t1, t2 UNODC Workshop, November, Bogota, Colombia 29 Good input data/model and stable coverage Stable forest, Stable mature coca, Stable young coca, Stable bare soil, BuildRoad, Stable other Good input data/model and dynamic coverage Cropped coca, Growing mature coca, New plots in the forest, New plots with young coca Sufficient but partially contradictory input data/model Forest in any time, Suspect coca in any time, Young coca in any time, Bare soil in any time Insufficient input data/model ProbBareSoil, ProbForest, Always clouds or shadows, sub_instableresidual

30 Mapping Out UNODC Workshop, November, Bogota, Colombia 30

31 Accuracy time0 UNODC Workshop, November, Bogota, Colombia 31 Table 1 Accuracy Assessment: Best Class at Time 0 (January 2004) User Class \ Mature Forest Bare Young Sum Sample Coca Soil Coca Confusion Matrix Mature Coca Forest Bare Soil Young Coca Unclassified Sum Accuracy Producer User Hellden Short KIA Per Class Totals Overall Accuracy KIA

32 Accuracy Time1 UNODC Workshop, November, Bogota, Colombia 32 Table 1 Accuracy Assessment: Best Class at Time 1 (March 2004) User Class \ Mature Forest Bare Young Sum Sample Coca Soil Coca Confusion Matrix Mature Coca Forest Bare Soil Young Coca Unclassified Sum Accuracy Producer User Hellden Short KIA Per Class Totals Overall Accuracy KIA

33 Accuracy Time2 UNODC Workshop, November, Bogota, Colombia 33 Table 1 Accuracy Assessment: Best Class at Time 2 (June 2004) User Class \ Mature Forest Bare Young Sum Sample Coca Soil Coca Confusion Matrix Mature Coca Forest Bare Soil Young Coca Unclassified Sum Accuracy Producer User Hellden Short KIA Per Class Totals Overall Accuracy KIA

34 Average Accuracy UNODC Workshop, November, Bogota, Colombia 34 Table 1 Accuracy Assessment: Average Best Class on Time 0, 1, and 2 (January, March, and June 2004) General User Class \ Sample Mature Coca Forest Bare Soil Young Coca Sum Confusion Matrix Mature Coca Forest Bare Soil Young Coca Unclassified Sum Accuracy Producer User Hellden Short KIA Per Class Totals Overall Accuracy KIA

35 Statistics UNODC Workshop, November, Bogota, Colombia 35 Code CLASS Number of Objects Surface (sqkm) Surface % 1st MBR Average 2nd MBR Average Classification Stability Average Input Data/Model Quality Surface (sqkm) Surface % 11 Stable forest % Stable mature coca % Stable young coca % Stable bare soil % BuildRoad % Stable other % Good Static % 21 Cropped coca % Growing mature coca % New plots in the forest % Good 24 New plots with young coca % Dynamic % 31 Forest in any time % Suspect coca in any time % Young coca in any time % Bare soil in any time % Sufficient but partially contradictory % 91 ProbBareSoil % ProbForest % Always clouds or shadows % sub_instableresidual % Insufficient % Grand Total %

36 Conclusions UNODC Workshop, November, Bogota, Colombia 36 VHR data can be suitable for automatic recognition of coca fields Optical tested Radar to be tested New approach needed Exploitation of image textural information Multi-criteria fuzzy inferential engine Multi-temporal improve reliability Possible uses Improve consistency of visual interpretation Automatise reference/sample processing Constraints 1 meter resolution at least Optical = cloud cover problems JRC is available for methodological support to UNODC More tests, fine-tuning in select test sites Test new sensors (VHR radar)

37 Thanks UNODC Workshop, November, Bogota, Colombia 37 Martino Pesaresi European Commission, Joint Research Centre, Institute for the Protection and Security of the Citizen (IPSC) Via E. Fermi, 1 TP 267 Ispra (VA), Italy Tel (direct line) Fax martino.pesaresi@jrc.it Web: