Landsat/TMによる河川境界検出に対する多重Morphology解析の応用

抄録

In a classification of satellite remote sensing image data, spectral distribution in a feature space is usually used. Mixels caused by class boundaries, however, give miss-classification results. To solve this problem many researchers have tried to extract or to avoid mixels using edge detectors such as spatial frequency filters. In this study, we describes a method to make clear edge information mentioned above. In particular, a new edge detector using Mathematical Morphology algorithm is introduced.
Conventional Morphological edge detectors are classified in dilation-type and erosion-type. The dilation-type edge detector is defined by subtraction of the original image from the dilated image. In a similar manner, the erosion-type edge detector is defined by subtraction of the eroded image from the original image. Both dilationtype and erosion-type edge detectors can detect edge information clearly. Their outputs, however, show different gradients depending on the formula and the size of structuring element.
From this point of view, a Multi-functional Wide-Narrow Morphological Edge detector (WNED) is introduced. WNED consists of wide range edge detection (WED), narrow range edge detection (NED) and their subtraction. At the first step of WNED, the maximization of dilation- and erosion-type NED are calculated. WED is then calculated. Finally the subtraction of them is calculated.
Performance of WNED is shown by applying visible and IR image of Landsat TM data which cover Shonai river and Kiso river, Japan and compared with the results analyzed by conventional formulae. As the results of case study, followings are found. Application of dilation- and erosion-type edge detectors is not sufficient to pick up the weak gradients. Even if the size of structuring element for dilation- and erosion-type ones is increased, undesireble gradients are detected. In contrast to such algorithms, WEND can detect the desired gradients and can delete the spurious edge pixels by subtracting NED gradients from WED gradients.