Tree crops and agroforestry systems are very representative of tropical agricultural landscape. Automatic recognition and mapping of this typical land cover type is thus a challenge for the use of remote sensing in driving issues in food sustainability. Therefore, this paper presents an attempt to use the potential of the object?based approach in image classification to produce a land?cover map of a complex agroforestry area. This case?study focuses on very high spatial resolution data acquired over the savannah/cocoa/forest transition region of Bokito in Cameroon, providing a large panel of various cropping systems. WorldView2 panchromatic and multispectral data are thus processed through textural indices derivation and principal component analysis, to select the more discriminant attributes for the different land?cover types, resulting in a stack of 32 image layers. The object?based approach is then run on eCognition Developer, combining several levels of multiresolution segmentation and consecutive classifications that involves different criteria at each step. At the end, a land?use map based on 13 classes was produced with 85% of global accuracy, evaluated based on ground?truth data and photointerpretation. Its typology is rather fine, especially for the agroforestry crops displaying complex structures, and that would not have been accurately delimitated nor discriminated with a pixel?based approach.