Edge detection as a starting point for remote sensing scene interpretation

We propose an edge-constrained smoothing filter to reduce within-class variance prior to spectral classification of remote sensing imagery. This is motivated by the scale mismatch between certain image resolutions and the landcover objects represented, leading to violation of an assumption on which pixelwise spectral classification is based. The first step of our procedure is edge detection within the scene, and then smoothing takes place largely within the confines of those edges, thus introducing spatial context to the smoothed value of any pixel. Our filter is adaptive to local variance so that the high variance areas undergo a high level of smoothing, while other areas are modified less. The smoothed image is processed with a conventional supervised classification, and the results compared to thematic maps obtained without the smoothing step. Comparisons using SPOT imagery of a complex urban and peri-urban environment show that our method decreases within-class variance and increases class separability, leading to landcover polygons that are relatively clean and useful as GIS objects