Paddock segmentation using multi-temporal satellite imagery

In previous work we demonstrated the use of temporal image sequences to identify broad land use classes [1, 2], relying on analysis at paddock rather than pixel scales. This paper describes a method to spatially segment a time-series stack of imagery to produce paddock units. Results are demonstrated on a significant portion of the Canterbury Plains where over 58,000 paddocks were identified. This enables regional-scale analysis of agricultural patterns in this rapidly changing landscape. Visually, the results appear in excellent agreement with the imagery. Analysis of linework around paddocks shows a mean deviation of 5.62 m (around half the imagery's pixel size), compared to hand-drawn maps. More difficulty was encountered in finding suitable methodology to assess accuracy at the paddock level. Initial results show the more critical under-segmentation of paddocks occurring 22% of the time. However, much of this error could be due to a 3-year gap from the reference imagery, or the occurrence of adjacent paddocks with identical crops.