A COMPARATIVE ASSESSMENT OF SUPERVISED PIXEL–BASED CLASSIFICATION METHODS IN THE DETECTION OF LANDSLIDE SCARS

M. LOUSADA1, C. LIRA1, P. PINA1, A. GONÇALVES2, A. P. FALCÃO2, S. HELENO2, M. MATIAS2, A. J. DE SOUSA1, M. J. PEREIRA1, R. OLIVEIRA3 AND A. B. ALMEIDA3

1CERENA, 2ICIST, 3CEHiDRO

IST/UTL - INSTITUTO SUPERIOR TÉCNICO, LISBOA, PORTUGAL

INTRODUCTION

LANDSLIDE SCAR LANDSLIDE TRACK GRASS BARE SOIL CLOUDS GRAVEL FOREST ROOFS INDUSTRY SHADOW ROADS

On February 20th, 2010 heavy rainfall culminated in violent floods and mudslides in Madeira Island, Portugal. This extreme event triggered thousands of landslides in both inhabited and uninhabited zones, resulting in extensive personal and material damages. Two main areas were heavily affected: Funchal and Ribeira Brava. Aiming to estimate the volume of sediment displaced during the event, a landslide inventory was an urgent necessity. This paper focuses on the procedures used in the cartographic inventory of the landslides, particularly in the assessment tests for the most accurate automatic classification procedure and the applied post-processing methods. GeoEye-1 satellite imagery from February 23rd and 28th, 2010, was the basis for our classification procedures.

RESULTS Maximum Likelihood

Overall Accuracy = (969054/1033569) 93.76% Kappa Coefficient = 0.915

Mahalanobis Distance

Overall Accuracy = (836836/1033569) 80.97% Kappa Coefficient = 0.749

Parallelepiped

Overall Accuracy = (336155/1033569) 32.52% Kappa Coefficient = 0.243

Minimum Distance

Overall Accuracy = (831048/1033569) 80.41% Kappa Coefficient = 0.742

CLASSIFICATION

A)

CLASSIFICATIONS IN GEOEYE -1 IMAGERY

(R-G-B-NIR band, 2 m/pixel and panchromatic 0.5 m/pixel

resolution)

CONFUSION MATRICES WITH GROUND TRUTH

IMAGE BUILT FROM ROIS

POST-CLASSIFICATION OF MAXIMUM LIKELIHOOD

SIEVING AND CLUMPING

MANUAL DELINEATION AND VALIDATION OF

LANDSLIDE SCARS WITH ORTOPHOTOMAPS

B) Computation of the confusion matrices allowed the comparison of results and the evaluation of the accuracy of landslide scar classification with a ground truth image built from ROIs.

Several training areas or regions of interest (ROIs) were selected, different sets of ROIs were tested and the best were used to create a total of 11 classes . Results indicated that the Maximum Likelihood algorithm presents the best accuracy and quality on landslide scar contours.

D) To correct contours and obtain the eventually missing scars , manual delineation and corrections were edited, with high resolution ortophotomaps (1:5000 scale with 0.4 m ) from may, 2010.

C) In a post-classification stage tools as sieve, clump and majority analysis were used in the Maximum Likelihood classification and tested with different thresholds to suppress or clump isolated or small groups of pixels, improving widely the quality of the final landslide scar layout.

The final results are very satisfactory, as the methodology produced an overall accuracy over 90% in the detection of landslides for the study area. This enabled an extensive inventory of the scars, with a substantially less time-consuming and less expensive process than the traditional manual delimitation methods.

CONCLUSION

Clump Sieve

Maximum Likelihood GeoEye image Mahalanobis Distance (pan-sharpening 0.5 m/pixel)

Minimum Distance GeoEye image Parallelepiped (pan-sharpening 0.5 m/pixel)

Landslide Scars, polygons manually corrected/validated

Landslide Scars, polygons obtained from Max-Like classification

Ortophotomaps with 0.4m spatial resolution

Maximum Likelihood classification

A)

METHODS

Ground Truth (Percent)

Class Lnds. scar Lnds. track Grass Bare soil Clouds Gravel Forest Roofs Industry Shadow Roads Total

Landslide scar 92,58 3,81 0 0,91 0 0 0,06 1,03 0,07 0 0,12 0,6

Landslide track 3,26 95,91 0 0,01 0 0 0 0,67 0 0 0 0,71

Grass 0,52 0,05 99,82 0 0 0,01 0,35 0 0 0 0 0,89

Bare soil 2,4 0,09 0 99,08 0 0 0,02 0 0 0 0 0,84

Clouds 0,13 0,07 0 0 99,01 0,07 0,04 0 0,19 0 0,32 26,32

Gravel 0,12 0,03 0 0 0 58,01 0,15 2,22 15,77 0,48 7,22 3,81

Forest 0,3 0 0,18 0 0 0,47 93,39 0,01 0,1 2,29 0,11 31,8

Roofs 0,69 0,04 0 0 0 0,01 0 93,23 0,31 0 0,08 0,66

Industry 0 0 0 0 0,99 3,35 0 1,4 80,41 0 1,14 2,82

Shadow 0 0 0 0 0 2,27 5,99 0 0,03 97,22 0 28,87

Roads 0 0 0 0 0 35,8 0 1,43 3,13 0 91,01 2,68

Total 100 100 100 100 100 100 100 100 100 100 100 100

Commission Omission Commission Omission User Acc. Prod. Acc. Prod. Accuracy User Accuracy

Class (Percent) (Percent) (Pixels) (Pixels) (Percent) (Percent) (Pixels) (Pixels)

Landslide scar 11,17 7,42 693/6206 442/5955 92,58 88,83 5513/5955 5513/6206

Landslide track 3,31 4,09 244/7366 304/7426 95,91 96,69 7122/7426 7122/7366

Grass 13,59 0,18 1246/9168 14/7936 99,82 86,41 7922/7936 7922/9168

Bare soil 2,36 0,92 204/8634 78/8508 99,08 97,64 8430/8508 8430/8634

Clouds 0,1 0,99 283/272044 2709/274470 99,01 99,9 271761/274470 271761/272044

Gravel 18,76 41,99 7390/39395 23164/55169 58,01 81,24 32005/55169 32005/39395

Forest 2,08 6,61 6841/328725 22797/344681 93,39 97,92 321884/344681 321884/328725

Roofs 2,15 6,77 146/6782 482/7118 93,23 97,85 6636/7118 6636/6782

Industry 16,25 19,59 4745/29192 5955/30402 80,41 83,75 24447/30402 24447/29192

Shadow 7,34 2,78 21917/298406 7894/284383 97,22 92,66 276489/284383 276489/298406

Roads 75,25 8,99 20806/27651 676/7521 91,01 24,75 6845/7521 6845/27651

GeoEye image ( 0.5m/pixel)

B) C) D)