Image Enhancement and Feature Extraction Based on Low-Resolution Satellite Data
The purpose of this study is to investigate the sensitivity of contrast-based textural measurements and morphological characteristics that derive from high-resolution satellite imagery (three-band SPOT-5) when diverse image enhancements techniques are piloted. The general framework of the application is the built-up/nonbuilt-up detection. In the existence of a low-resolution reference layer, we apply supervised learning that indirectly reduces the uncertainty and improves the quality of the reference layer. Based on the new class label assignments, the image histogram is adjusted suitably for the computation of contrast-based textural/morphological features. A case study is presented where we test a mixture of image enhancement operations like linear and decorrelation stretching and assess the performance through ROC analysis against available building footprints. Experimental results demonstrate that spectral band combination is the key factor that conditions the contrast of grayscale images. Contrast adjustment (before or after the band combination and merging) supports considerably the extraction of informative features from a low-contrast image; in case of a well-contrasted image, the improvement is marginal.
Built-up detection, contrast adjustment, feature extraction, high-resolution image enhancement, low-resolution reference data, morphological, supervised learning, support vector machines (SVMs), textural.