High resolution remote sensing imagery is becoming a reality in the context of Precision Agriculture. In future, it will be an essential tool for site specific crop management and will help in the global knowledge of land use and of "terroir" quality. In the Bordeaux wine producing area, France, all the parties involved in wine making (vine growers, cooperatives, vine grower unions, etc.) are interested in a better knowledge of vineyards. They are expecting precision viticulture technologies and more particularly remote sensing imagery to provide them with detailed maps (e.g. vegetative vigor maps) of vine fields. Such maps can be frequently updated and enable the determination of management zones inside the vine fields. Various processing steps are involved in the design of management maps, going from image acquisition to the integration into a GIS. In a previous paper, we addressed the accurate detection of vine rows inside the field using a network of deformable templates. In the present paper, we focus on a preliminary processing step, the automatic delineation of a specific field within an image. The objective of this step is the automatic extraction of the field of interest from its neighborhood (roads, buildings, trees and other fields). Automatic vine field delineation deals with segmentation algorithms. Indeed, segmentation algorithms have been widely used in the framework of remote sensing. Such algorithms can be classified into three groups: split and merge algorithms, statistical approaches and region growing methods. The split and merge algorithms are the most common segmentation methods. They proceed in two steps. Firstly, they split iteratively the image, until each region is homogeneous. The resulting set of region is often over-segmented. A quad tree algorithm usually fulfils such a task. A second step is then necessary to merge neighboring regions providing that they are similar. These methods are easy to implement, but they tend to favor rectangular regions. Moreover, they are very sensitive to the threshold values used to determine the homogeneity and the similarity of the regions. The statistical approaches are based on a joint statistical modelization of the region shape and of their content. Among these approaches, the markovian models are the most frequently used, thanks to their flexibility and their efficiency. The only drawback of such statistical approaches is the huge computation time needed to guaranty an optimal segmentation. Sub-optimal algorithms exist but their reliability depends on their initialization. In this paper we choose a segmentation approach derived from the region growing method. These methods appeared in the early 1970's. They operate from seed pixels belonging to the region to be segmented. Then, pixels are added iteratively to this initial region provided that the candidate pixel shows similar features than those of the initial region. The algorithm stops when no pixel can be added to the region. This class of methods is specifically suitable in the case of the delineation of a unique region in a picture, but it requires some supervision from the operator in order to choose the initial germ inside the region. We consider here that this very simple task is acceptable. The paper present the acquisition method and images, the extraction of the textural features used in the segmentation algorithm, and the region growing algorithm.