Feature extraction for high-resolution imagery based on human visual perception

Feature extraction is highly important for classification of remote-sensing (RS) images. However, extraction of comprehensive spatial features from high-resolution imagery is still challenging, leading to many misclassifications in various applications. To address the problem, a shape-adaptive neighbourhood (SAN) technique is presented based on human visual perception. The SAN technique is an adaptive feature-extraction method that not only considers spectral feature information but also the spatial neighbourhood as well as the shape of features. The distinct advantage of this approach is that it can be adjusted to different feature sizes and shapes. Assessment experiments on a Système Pour l'Observation de la Terre 5 (SPOT-5) image were conducted to perform classification of land use/land cover. Results showed that improvement with SAN features is not significant for supervised classifiers due to the spectral confusion problem that resulted from similar spectral signatures between farmland and green areas, but a particularly significant improvement is observed for the unsupervised classifier. For the unsupervised classification, the SAN features noticeably improved the overall accuracy from 0.58 to 0.86, and the kappa coefficient from 0.45 to 0.80, indicating promise in the application of SAN features in the auto-interpretation of RS images.