Remote Sensing and Geoinformation

The aim of this research is to determine an accuracy assessment of different multispectral gradient-based edge detectors. We will present and evaluate three different approaches: the mean-, the maximumand the multispectral gradient approach. The mean approach determines the overall gradient as the arithmetic mean of all (band-wise) gradient vectors, whereas the maximum approach selects the gradient vector of maximum length. The first two algorithms that are heuristically motivated, the multispectral gradient approach can be derived mathematically from the single band gradient-based approach and thus is very interesting to investigate (see [1]). To compare and evaluate the algorithms, we designed a modular framework that is based on generic programming and the VIGRA computer vision library [2]. We discuss the framework's architecture in more detail to demonstrate the flexibility. For the evaluation we synthesized artificial images where we know the exact location and occurrence of edge elements (evaluation by means of dedicated techniques [3]). The evaluation shows that in many cases the naive mean approach does not lead to satisfactory results. In some cases, the maximumand the multispectral gradient approach are almost on the same level of detection quality. In other cases the multispectral gradient approach outperforms the other two approaches. Complementary to the quantitative evaluation, we present the application of the algorithms to Landsat 7 ETM+ satellite multispectral imagery of coastal and urban areas taken from the public Landsat Archive.