Identification of orchards and vineyards with different texture-based measurements by using an object-oriented classification approach

This article presents an object-oriented classification approach to identifying orchards, vineyards and agricultural fields. This approach uses texture-related parameters obtained from very high spatial resolution data, in particular Quickbird images and orthophotos. A multi-resolution segmentation procedure was applied to delimit individual agricultural plots as segments. Textural information of the generated segments was then used to classify orchards, agricultural fields (grassland) and two wine cultivation systems (trellis and pergola). In this article three different texture-based approaches are compared to correctly classify the given plots: (1) a 'zonal mean maximum' of texture measurements, which consider the maximum value of four directions of texture measurements related to plots; (2) a relational sub-object approach based on a thematic derived texture filter technique that reflects individual row structures; and (3) a hybrid approach combining the two previous ones. In order to identify relevant parameters for each approach, the data mining software See5 is used. The hybrid approach increased overall accuracy by 8% and 6% for Quickbird (92% accuracy) and orthophotos (88% accuracy), respectively. The application of the same methodology to the orthophotos alone results in a lower accuracy but still one that is acceptable. This offers the possibility of also considering orthophotos for this kind of detection, especially when Quickbird data are not available. In this sense, the developed methodology can be considered as a new object-based landscape analysis technique suitable for the provision of accurate maps able to fulfil the requirements of scientists, planners and other end-users.