Boundary Delineation of Agricultural Fields in Multitemporal Satellite Imagery

Abstract

Agricultural land-use statistics are more informative per-field than per-pixel. Land-use classification requires up-to-date field boundary maps potentially covering large areas containing thousands of farms. This kind of map is usually difficult to obtain. We have developed a new, automated method for deriving closed polygons around fields from time-series satellite imagery. We have been using this method operationally in New Zealand to map whole districts using imagery from several satellite sensors, with little need to vary parameters. Our method looks for boundaries—either step edges or linear features—surrounding regions of low variability throughout the time series. Local standard deviations from all image dates are combined, and the result is convolved with a series of extended directional edge filters. We propose that edge linearity over a long distance is a more important criterion than spectral difference for separating fields, so edge responses are thresholded primarily by length rather than strength. The resulting raster edge map (combined from all directions) is converted to vector (GIS) format and the final polygon topology is built. The method successfully segments parcels containing different crops and pasture, as well as those separated by boundaries such as roads and hedgerows. Here we describe the technique and demonstrate it for an agricultural study site (4000 km2) using SPOT satellite imagery. We show that our result compares favorably with that from existing segmentation methods in terms of both quantitative quality metrics and suitability for land-use classification.