by Megan Kanaga Creutzburg

Feature Extraction

Target detection refers to the use of high spectral resolution remotely sensed images to map the locations of a target or feature (often a plant species of interest) with a particular spectral or spatial signature. Target detection or feature extraction encompasses a broad range of techniques, including measurements derived from individual bands (eg. thermal anomalies) and more complex methods designed to recognize discrete features by shape, hyperspectral signature, or texture (see texture analysis). Targets of interest are often smaller than the pixel size of the image (subpixel target detection) or are mixed with other nontarget cover types within a pixel, requiring techniques such as spectral mixture analysis to detect the target species. Hyperspectral images are useful in target detection because they contain a large contiguous set of spectral bands, often numbering in the hundreds to thousands, and provide large quantities of high spectral resolution data. Using a hyperspectral image, the spectral properties of the target, such as contrast, variability, similarity and discriminability, can be used to detect targets at the subpixel level. The user specifies spectral endmembers, which are the reflectance spectra of the “pure” targets that occur across the landscape, and image processing software is used to characterize the extent of the target across the landscape. The selection of spectral endmembers is similar to the idea of identifying training areas in supervised classification, but the spectral endmember can then be used at a subpixel level to detect the species of interest. Spectral endmembers are often generated in the field using a field spectroradiometer. Then the image is processed using classification algorithms to detect the locations of the target species.

• Spectral mixture analysis
• Hyperspectral image classification
• Target detection can be similar to supervised and unsupervised classification, but is detecting a particular targeted vegetation type against a background of other vegetation, rather than classifying all vegetation
• Fractional cover is also determined using subpixel techniques
• Target detection can be used for purposes of change detection
• Boundary mapping

The output from target detection techniques is a map of the spatial distribution of the target object, species or cover type. Using subpixel techniques, the software estimates the abundance fractions of targets contained in each image pixel, rather than simply labeling each pixel to one cover class as in classical image processing.

Target detection techniques are generally used to map the presence of a species of interest, often invasive plants. Some specific rangeland uses are:

• Hamada et al. (2007) used target detection to map spread of invasive Tamarisk species in California
• Lass et al. (2005) mapped spotted knapweed and babysbreath invasions with target detection techniques
• Mundt et al. (2006) used multiple methods of target detection to track the invasive species Hoary cress and Rush skeletonweed in western Idaho
• Mundt et al. (2007) used target detection methods to map leafy spurge invasions in eastern Idaho

• Hamada, Y., D.A. Stow, L.L. Coulter, J.C. Jafolla, and L.W. Hendricks. 2007. Detecting Tamarisk species (Tamarix spp.) in riparian habitats of Southern California using high spatial resolution hyperspectral imagery. Remote Sensing of Environment 109: 237-248.
• Lass, L.W., T.S. Prather, N.F. Glenn, K.T. Weber, J.T. Mundt, and J. Pettingill. 2005. A review of remote sensing of invasive weeds and example of the early detection of spotted knapweed (Centaurea maculosa) and babysbreath (Gypsophila paniculata) with a hyperspectral sensor. Weed Science 53: 242-251.
• Mundt, J.T., N.F. Glenn, K.T. Weber, and J.A. Pettingill. 2006. Determining target detection limits and accuracy delineation using an incremental technique. Remote Sensing of Environment 205: 34-40.
• Mundt, J.T., D.R. Streutker, and N.F. Glenn. 2007. Partial unmixing of hyperspectral imagery: theory and methods. ASPRS 2007 Annual Conference. Tampa, Florida, May 7-11, 2007.

• Chang, C-I. and D.C. Heinz. 2000. Constrained subpixel target detection for remotely sensed imagery. IEEE Transactions on Geoscience and Remote Sensing 38: 1144-1159.
• Chang, C-I. and H. Ren. 2000. An Experiment-Based Quantitative and Comparative Analysis of Target Detection and Image Classification Algorithms for Hyperspectral Imagery. IEEE Transactions on Geoscience and Remote Sensing 38: 1044-1063.
• Geng, X.R. and Zhao, Y.C. 2007. Principle of small target detection for hyperspectral imagery. Sci China Ser D-Earth Sci 50: 1225-1231.
• Smith, A.M.S., M.J. Wooster, A.K. Powell, and D. Usher. 2002. Texture based feature extraction: application to burn scar detection in Earth Observation Imagery. International Journal of Remote Sensing. 23: 1733-1739.
• Tian, Y., P. Guo, and M.R. Lyu. 2005. Comparative studies on feature extraction methods for multispectral remote sensing image classification. 2005 IEEE International Conference on Systems, Man and Cybernetics. Waikoloa, Hawaii, October 10-12, 2005.
• Weber, K.T., J. Theau, and K. Serr. 2008. Effect of coregistration error on patchy target detection using high-resolution imagery. Remote Sensing of Environment 112: 845-850.

This method varies in accuracy based on the quality of the image, spatial and spectral resolution, and the degree of differentiation of the target spectral signature from the image background. Field data should be collected to verify remotely sensed maps of target species.

The requirements of the input image vary depending on the method of target detection or feature extraction. A multispectral (a set of multiple spectral bands) or hyperspectral (a large set of contiguous spectral bands) image obtained from a satellite or aircraft is often used to map the ranges of target species. The input image must be processed to correct for atmospheric effects (if derived from a satellite image) and any other necessary corrections. Vegetation indices are often used to aid in target detection, and other processing of the image based on the spectral properties of the target species may be necessary.

Target detection and feature extraction techniques require image processing software from companies such as ERDAS or ENVI.

The left image shows an n-dimensional visualization of the leafy spurge endmember (green), mixed pixels containing some leafy spurge (red) and spectrally similar vegetation that is not leafy spurge (blue). The right image shows the geographic distribution of each of the groups on a map (source: Mundt et al. 2007).



A hyperspectral image showing spotted knapweed (red) and babysbreath (green) distribution across sites in eastern Idaho (source: Lass et al. 2005).

_{You must have an account and be logged in to post or reply to the discussion topics below. Click here to login or register for the site.}