The objective of this work is to verify the viability of cocoa agroforestry classification with High Resolution imagery to include, in the cropland area, the mapping of cocoa planted under forest, as well as open cocoa plantation. In order to avoid overestimating the cocoa area, we introduce the concept of counter-examples (CE). Counter-examples are areas of known classes, not directly involved in classification focus, but identified to avoid the classes of focus being misleadingly classified. Two set of CE was used. The first one is the merging of 7 non-cocoa classes in one training set. The other uses each of these 7 CE classes separately in the training set. Among the several classifiers tested, the best one was SVM with RBF kernel. Results showed that using one CE set produces a more uniform classification map than using 7 CE separately and captures 20% more cocoa cultivated area in a test field, than mapping open cocoa only, with similar user accuracy.