Block-based semantic classification of high-resolution multispectral aerial images

In this paper, we compare different approaches for classification of aerial images based on descriptors computed using visible spectral bands as well as additional information obtained from the near infrared band. We also propose different methods for incorporating dimensionality reduction into descriptor extraction process for both global and local texture descriptors aiming at obtaining low-dimensional descriptors from multispectral images. Furthermore, we examine classification accuracy in cases when small training sets are used. For evaluation purposes, we use an in-house high-resolution aerial image dataset, with images containing visual and near-infrared spectral bands, as well as UC Merced land-use dataset. We achieve the classification rates of over 90% on in-house dataset. For UC Merced, we obtain classification accuracy of 91% which is an improvement of about 3% compared to the state-of-the-art color SIFT descriptors.