EXAMPLE OF OBJECT BASED IMAGE ANALYSIS APPLIED TO
COARSE RESOLUTION IMAGES. APPLICATION TO LANSCAPE MAPPING IN FRANCE
OBJECTIVE
Defining an objective methodology for obtaining a segmentation of landscapes over France using only satellite images. In this context ‘landscape’ is define as a ‘radiometrically homogeneous region’.
DATA
•Enhanced Vegetation Index (EVI) images:
•12 monthly images (average of 5 years)
•Texture images (Homogeneity and Entropy)
•4 monthly images (April, May, August, November)
RESULTS
Mar Bisquert(1), Michel Deshayes(1), Agnès Begué (2) (1)IRSTEA, Maison de la Télédétection, Montpellier
mar.bisquert-perles@teledetection.fr, michel.deshayes@teledetection.fr
(2)CIRAD, Maison de la Télédétection, Montpellier agnes.begue@teledetection.fr
CONCLUSIONS
-OBIA using coarse resolution images is a good instrument for delimiting the landscapes at regional scale.
-Texture indices add important information to vegetation indices, leading to a better stratification of landscapes.
-This segmentation of landscapes has to be linked to environmental variables (climate, topography, cartography of ecosystems,…) and to human activities to lead a landscape cartography.
REFERENCES
Johnson, B., Xie, Z., 2011. Unsupervised image segmentation evaluation and refinement using a multi-scale approach. ISPRS Journal of Photogrammetry and Remote Sensing 66, 473–483. Zhang, X., Xiao, P., Feng, X., 2012. An Unsupervised Evaluation Method for Remotely Sensed Imagery Segmentation. Geoscience and Remote Sensing Letters, IEEE 9, 156–160.
ABSTRACT
Object based image analysis (OBIA) is a powerful technique for stratification of remote sensing images. It has been widely used with very high resolution images, but it can be very useful also for classifying coarse resolution images, such as MODIS, Meris, and the future Sentinel-3 sensor, OLCI. In this study, an object based stratification of the French landscapes is made from MODIS images from years 2007-2011, taking into account spectral information (vegetation indices), temporal information (monthly images) and texture information. A non supervised classification is performed in order to obtain a classification of radiometrically homogeneous landscapes in France.
Segmentation with eCognition (Object Based image Classification) using different
combinations of input variables and different parameters of the segmentation method. METHODOLOGY
Comparison of the resulting segmentations and selection of the best one based on the objective evaluation methods proposed in Johnson et al. (2010) et Zhang et al. (2012) which measure the homogeneity of the regions and the disparity between regions. Johnson:
(
)(
)
norm norm j i uj n i i n i n j j i ij n i i n i i i M V J w y y y y y y w n M v v a V + = ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ − − − = =∑
∑
∑∑
∑
∑
≠ = = = = = 1 2 1 1 1 1 ) ( Zhang: ( )(max min) ( max min)
1 1 1 , log 1 10 1 D D T T D T Z n n mm m D a e n S T n b n i b ib n i i i − − = + = ⎟ ⎠ ⎞ ⎜ ⎝ ⎛ − = ⎟⎟ ⎠ ⎞ ⎜⎜ ⎝ ⎛ + =
∑∑
∑
= = = λ λn: total number of regions
viand ai: variance and area of segment I wij: a measure of the spatial proximity (is
equal to 1 if the regions are neighbours and 0 if not)
yi: mean spectral value of segment i
Y: mean spectral value of the image
Vnormand Mnorm: the normalised values of V
and M.
S: image size
ei: is the Euclidean distance between each
pixel and the average of region i.
mib: average value of band b in region i
mmb: average value of all regions in band b Tmax, Tmin, Dmaxand Dmin: are the maximum
and minimum values of T and D over a
set of segmentations obtained with
different homogeneity criterions.
Unsupervised
classification using the medium values of the input variables over the regions obtained from the best segmentation.
Analysis of the segmentation obtained using the 12 EVI images
Analysis of the segmentation obtained using the 12 EVI images and two textural indices (homogeneity and entropy)
Segmentation obtained with the 12 EVI images and the 8 images of homogeneity and entropy using 40 as scale parameter. 84 segments obtained.
RGB composition: R: EVI April, G: EVI June, B: EVI August
Entropy Homogeneity
EVI
Result of the unsupervised classification.