Application of wavelet transform for extracting edges of paddy fields from remotely sensed images

Abstract

For geographical information systems (GIS) to be useful in the management of Japanese paddy fields, it is vital to be able to prepare base maps that define sections of field. We attempted to prepare a base map by detecting the edges of submerged paddy fields from a SPOT image with the use of multi-resolution wavelet transform. The raw image of SPOT band 3 was decomposed into a low frequency approximation image and a set of high frequency detailed images to five levels, and then reconstructed. The reconstructed image at the fifth level was obtained from all of the high frequency images except for the low frequency approximation image. The image reconstructed up to full scale was then applied to a zero-crossing scheme and three post-processings—line thinning, removal of isolated pixels, and connection of pixels with their neighbours—in order to obtain clear edges. The spatial features of the image indicating the edges obtained by multi-resolution wavelet transform were compared quantitatively with those obtained by using a difference of Gaussian (DOG) filter. The multi-resolution wavelet transform was better than the DOG filter in that the base map obtained by the wavelet transform method represented a ‘field block’, which is typically composed of 10 holdings of paddy field, more precisely than with the DOG filter, and in that the number of open polygons on this map was smaller, indicating that the wavelet transform method is more suitable for producing the base map in GIS for the management of paddy fields.