In precision spraying, spray application efficiency depends on the pesticide application method, the phytosanitary product as well as the leaf surface properties. For environmental and economic reasons, the global trend is to reduce the pesticide application rate of the few approved active substances. Under these constraints, one of the challenges is to improve the efficiency of pesticide application. Different parameters can influence on pesticide application as nozzle types, liquid viscosity and leaf surface. Specific models have been developed showing that the predominant factor for the leaf is the leaf roughness, because it is related on adhesion mechanisms of liquid on surface and it is used to define nature of leaf (hydrophobic/hydrophylic). In this work, we focus on the leaf surface properties, in particular the vine leaves. We propose to discriminate between two kinds of vine leaves (Pinot and Chardonnay) for different stages of development to follow their roughness growth. Moreover, this discrimination has an impact on the product behavior, and allows adjusting the product viscosity and spraying parameters according to the roughness and the stage of vine leaf development. In the purpose of vine leaf discrimination, it is interesting to use texture analysis because of spatio frequential aspect of the features that can be extracted. In order to test the proposed leaf classification, experiments have been done on images acquired with a SEM microscope. These images represent various surfaces (above, below, with and without rib) of the vine leaves for different stages of development: "young" and "mature" leaves. Generally, the major issue with natural texture classification is the sensitivity to illumination, changes of scale and orientation. The overall performance of a texture classifier may be totally degraded if the unknown patterns to be classified are slightly rotated with respect to the training samples. In our case we consider the invariant features (scale, illumination and rotation) called Generalized Fourier Descriptors (GFD). The application of Generalized Fourier Descriptors allows the extraction of a vector robust texture features. However, this vector has high dimensionality which can cause erroneous classification due to the Hughes phenomenon. To avoid this constraint, it is interesting to apply dimensionality reduction (RD) techniques in order to obtain representative data with reduced size. Nonlinear RD technique as Kernel Discriminant Analysis (KDA) is used, which is the most commonly used and the most suitable for our application. A multi-layer perceptron will be used as classifier tool. The result shows that the combination of GFD and KDA appears to provide sufficient information to characterize vine leaves for different stages of development. The rate of classification for young leaves is 100%, and 94.44% for mature leaves of Chardonnay, 97.22% for young leaves and 98.55% for mature leaves of Pinot.