Segmenting multispectral Landsat TM images into field units

Abstract

Presents a procedure for the automated segmentation of multispectral Landsat TM images of farmland in Western Australia into field units. The segmentation procedure, named the canonically-guided region growing (CGRG) procedure, assumes that each field contains only one ground cover type and that the width of the minimum field of interest is known. The CGRG procedure segments images using a seeded region growing algorithm, but is novel in the method used to generate the internal field markers used as seeds. These internal field markers are obtained from a multiband, local canonical eigenvalue image. Before the local transformation is applied, the original image is morphologically filtered to estimate both between-field variation and within-field variation in the image. Local computation of the canonical variate transform, using a moving window sized to fit just inside the smallest field of interest, ensures that the between- and within-field spatial variations in each image band are accommodated. The eigenvalues of the local transform are then used to discriminate between an area completely inside a field or at a field boundary. The results obtained using CGRG and the methods of Lee (1997) and Tilton (1998) were numerically compared to ideal segmentations of a set of sample satellite images. The comparison indicates that the results of the CGRG are usually more accurate in terms of field boundary position and degree of over-segmentation and under-segmentation, than either of the other procedures.