|M.Sc Student||Braitbart Mor Ester|
|Subject||Multi-Spectral Edge Detection for Enhanced Extraction and|
Classification of Homogeneous Regions in Remotely
|Department||Department of Civil and Environmental Engineering||Supervisors||Professor Maxim Shoshany|
|Dr. Ophir Regev Almog|
|Full Thesis text - in Hebrew|
Multi-spectral images of Mediterranean environments are characterized by high spatial and temporal heterogeneity that originate from various factors: (1) The geo- diversity of the Mediterranean environment; (2) acquisition conditions; (3) medium spectral and spatial resolutions.
Mapping land covers under such heterogeneity is a challenging task. The objective of this study is to improve classification results for areas with high spatial and spectral heterogeneity by applying object-oriented classification instead of the common per pixel classification, following image objects separation through edge detection.
Image processing techniques applied for this purpose, including segmentation and classification, attracted significant attention during the past decades. Most segmentation approaches are single band, few are RGB and less multi-spectral, do not exhaust the wealth of spectral information in the image. These methods mainly apply the principle of minimum internal homogeneity and maximum difference between neighboring regions.
A multi spectral approach was developed for defining edges as strips which separate relatively homogenous image regions, representing transition zones characterized by mixtures of adjacent land covers . The main elements of our methodology are: (1) the calculation of the spatial variance at different bands and maximizing their combination in extracting edges ; (2) enhancement of edginess according to the image homogeneity around edge pixels; (3) per-object classification for the homogenous areas delineated between edge strips ; (4) adding information from ISODATA unsupervised classification to large image areas which were unclassified in the previous stage (3). This methodology was applied separately for vegetated and bare image areas based on the implementation of a NDVI threshold.
Our methodology was applied on multi-spectral images acquired by the VENμS remote sensing system (Vegetation and Environment on a New Micro Satellite). The satellite is equipped with a multi-spectral camera with 12 narrow spectral bands at 5m ground resolution. From the images that were calibrated to reduce atmospheric effects, test areas with different contours of urban areas, agricultural areas, roads and vegetation areas were analyzed.
For the quantitative evaluation of the results, the use of confusion matrix facilitated estimating the overall accuracy and the calculation of the entropy (Shannon Information) of the classified images allowed calculation of the degree of order/ disorder of the classified image (high proportion of small classified patches represents high disorder).
The methodology showed good performance separating agricultural parcels and relatively homogenous bare surface areas. In vegetated areas, the algorithm showed lower efficiency and was affected by the spatial variability resulting from different shading and vegetation combinations. A good result concerned the separation of built areas and roads from the other land cover types.