This document presents a methodology for land use mapping using segmentation techniques on coarse resolution SAR data. It explores extracting urban extents using the BuiltArea algorithm, then segmenting the SAR images using different algorithms like Canny edge detection. Land use classes like commercial, residential and green areas are classified after feature selection and majority rule application. Testing on Shanghai showed potential for moderate SAR in urban monitoring. Preliminary fusion with optical data from Beijing-1 satellite improved segmentation accuracy and classification results. Future work will explore polarimetric features and additional classes with multitemporal SAR segmentation approaches.
1. LAND-USE MAPPING USING COARSE RESOLUTION SAR DATA AT THE OBJECT LEVEL EXPLOITING ANCILLARY OPTICAL DATA ALDRIGHI M., GAMBA P. Remote Sensing Team UniversitàdegliStudi di Pavia Italy
2. OUTLINE Introduction Problem Analysis ProposedMethods SegmentationTechniques Exploitation of Ancillary Optical Data Case Study: ENVISAT/ASAR, Shanghai Conclusions
3. INTRODUCTION The work is subdivided into two different tasks: the first one is the extraction of the human settlement extents for different sensors; implementation of “BuiltArea” procedure the second one requires instead the characterization of different land use classes using the same SAR data sets. Different segmentation techniques are compared and exploited in order to identify statistically homogeneous regions and eventually a supervised classification of the selected features allows assigning each region to a class; Tests on the areas of Shanghai show potentials for the use of these techniques in urban area monitoring using moderate resolution SAR; A preliminary study on the exploitation of ancillary optical data for segmentation.
4.
5. We rely on the more precise version of the “Built-Area” algorithm
6. All of the most recent approaches developed for SAR images rely on spatial indexes and/or a combination to extract human settlements
7. The indexes considered in this work belong to two major categories: Local Indicators of Spatial Associations (LISA) and co-occurrence textural features:
12. Variance.Moran Geary BuiltArea algorithm Urban Areas Extraction Getis-Ord Density Analysis Morphological Operation Correlation Variance
13. BUILTAREA - Beijing COR 5 B.A. INPUTS VAR MORAN 3-PARAMETERS SET ACCORDING TO THE ACQUISITION SENSOR (ENVISAT/ASAR): SCALE_LISA: 0.6 SCALE_TEXT: 0.4 SCALE_URB: 0.1 GEARY GETIS-ORD ENVISAT/ASAR APP - Geocoded 12,5 m spatial resolution Acquired on August 8° 2009
14.
15. Urban environments show a natural structural organization, and they can be seen as block agglomerates rather than building units.
16. It makes sense to segment a SAR image into statistically homogeneous areas and use these regions as a spatial proxy to urban blocks.
20. valid for everysegmentorientationItisbased on fourdifferentphases: Image Smoothing, Edge Enhancement, Non-maximasuppression, HysteresisThresholding. PROCESSING CHAIN based on: a denoising algorithm, an edge detection step, a region merging technique
21. REGION MERGING REGION MERGING It uses the output of the canny edge detector in order to generate, from the original edge detected image, well-defined closed regions corresponding to statistically homogeneous areas.
22. ImageSeg Algorithm The Berkeley ImageSeg algorithm is an object-based image analyzer algorithm, where compactness, shape and scale parameters may be adjusted in order to obtain the desired level of segmentation. ASAR/ENVISAT APP IMAGE VV-POLARIZATION PIXEL POSTING: 12.5 m
23. Spanning Tree Based Algorithm The Marpu algorithm is based on a graph theoretic approach together with a region growing technique where the graph is used to guide the merging process. The algorithm is based on building a graph over the image connecting all the objects. The Standard deviation to Mean Ratio (SMR) is used as the homogeneity criterion while merging the objects. Higher value of this ratio will yield bigger objects and vice-versa.
24. PROCESSING CHAIN URBAN EXTENT EXTRACTION LISA AND TEXTURE COMPUTATION BUILTAREA APPLICATION SEGMENTATION CANNY EDGE DETECTOR ImgSeg Spanning Tree TRAINING SITES SELECTION SITES IDENTIFICATION HOMOGENIZATION – SEGMENTATION BASED FEATURES REDUCTION JEFFRIES-MATUSITA INDEX CLASSIFICATION SUPERVISED CLASSICIATION MAJORITY RULE APPLICATION
25. TRAINING SITES SELECTION Training set ismodifiedaccording to the segmentation. Only the twowidestregionsbelonging to it are maintained.
26. FEATURES SELECTION In order to maintain the number of featuresas small aspossible, the two more representativefeature are used for the classificationstep. MoranIndex GearyIndex Getis-OrdIndex Correlazione Jeffries – Matusita Index Varianza LISA Density OriginalValues SpeckleDivergence L.D.I.
27. MAJORITY RULE APPLICATION Regionscontourssuperimposition over classifiction. Re-classificaitionbased on majorityvoting per region. class1 class 2 class 3
33. SEGMENTATION CannyEdge Detector and RegionMerging A B C D Edge Detection EdgeMap RegionMerging ShapeFiles BIS Algorithm Spanning Tree Algorithm
34. IMGSEG VS SPANNING-TREE BASED Final Output of the processing chain, after the Minimum Distance classification and the application of the majority rule. LEGEND BLUE: water YELLOW: Commercial Areas RED: Residential C.D. IMGSEG – Final Output Spanning Tree - Final Output Ground Truth ASAR/ENVISAT Image
