Wavelet texture features for improved classification of quickbird imagery
The availability of high-resolution satellite imagery has rapidly increased over the last six years since the introduction of commercial satellite imagery such as Quickbird, IKONOS, and OrbView. A unique characteristic of these sensors is that they combine a very high-resolution ($<$~1~m) panchromatic band with four (lower resolution) multispectral bands. Many studies have focused on combining these five bands of different spatial resolutions to obtain a multispectral image at 0.6~m or 1.0~m resolution. This process is often referred to as pansharpening.
The emphasis of image processing techniques for these high-resolution images has shifted from traditional pixel-based techniques to contextual and object-oriented approaches as the size of the object of interest is often larger than one pixel. Many studies have explored the use of texture measures for improving classification or segmentation results by including the spatial domain. These texture measures are commonly based on a single image band and rarely combined with spectral information.
This study presents a novel approach in combining textural information from the panchromatic band with spectral information from the multispectral bands for improved image classification. Firstly, we develop a texture measure based on wavelet coefficients of the panchromatic band that can be aggregated to the resolution of the multispectral bands. Secondly, we combine the texture measures with the spectral information in the multispectral bands in a fuzzy classification framework. Thirdly, we illustrate our approach with a case study of vegetation and land cover classification based on a Quickbird image of sub-Antarctic Macquarie Island.
In our methodology, we work with a Quickbird image that contains four multispectral bands (B, G, R, NIR) at 2.4~m resolution and one panchromatic band at 0.6~m resolution. This means that each multispectral pixel contains 16 panchromatic pixels (4~by~4). We assume that the panchromatic band contains most information about the spatial characteristics or structural pattern of the land cover classes. The multispectral pixels on the other hand are only used for their spectral content. It should be stressed that the approach described here can be applied to other imagery that combine a high-resolution panchromatic band with multispectral bands.
In order to generate a measure describing the texture of the local neighbourhood of a multispectral image pixel, we define a neighbourhood of 16~by~16 panchromatic pixels centred around a multispectral pixel. A wavelet decomposition based on the Daubechies wavelet is then applied to this panchromatic image block. We assume that characteristic patterns at this image scale are depicted by high frequency changes in pixel reflectance. We therefore only use the small-scale wavelet coefficients to quantify texture. The structure of the wavelet coefficient matrix is effectively a measure for texture. To summarise this structure in a single texture measure we calculate the following statistics from the wavelet coefficient matrix: standard deviation, skewness, kurtosis, entropy, and energy. These measures are then assigned to the pixel at the resolution of the multispectral bands (2.4~m) at the centre of the 16~by~16 panchromatic pixel kernel. This approach effectively models the texture in the local neighbourhood of a multispectral pixel, based on the wavelet coefficients of the panchromatic pixels in this kernel. The statistics derived from the wavelet coefficients then quantify the texture for this area so they can be assigned as a collection of texture measures to the centre pixel. Combining the new texture image bands with the original multispectral bands results in an image stack containing spectral and spatial information that can be used in a subsequent land cover classification.
In our classification approach, we use a supervised variant of the fuzzy c-means classifier. This classifier relies on reference pixels in the image that represent specific land cover classes. The spectral and textural characteristics of these reference pixels are then used to train the classifier. The fuzzy classifier calculates the Euclidean distance to the 5 nearest reference pixels in each class in multi-dimensional feature space for all unlabeled pixels. These distance measures are used to compute fuzzy membership values for each pixel to every class, depicting the degree of similarity to each class. This approach is effectively a fuzzy k-nearest neighbour (kNN) algorithm. The main advantage of this approach is that we obtain a measure of classification uncertainty.
The algorithm is illustrated with a land cover classification based on a Quickbird image of Macquarie Island acquired in March 2005. Macquarie Island is a sub-Antarctic island with unique vegetation communities and geological features. It is a designated World Heritage Area. High-resolution satellite imagery provides an effective source of information to map and monitor the Island's unique sub-Antarctic vegetation, which is rapidly changing under the pressures of rabbit grazing. Some vegetation types, such as Tussock grass, show characteristic patterns on the Quickbird image. With the lack of spectral resolution we argue that it is beneficial (if not crucial) to include information about the spatial characteristics in an image classification. With reliable and accurate field samples as reference data we show that the classification accuracy dramatically increases by taking texture measures into account, exploiting both the spatial and spectral information in Quickbird imagery.