The categorical description of leaf shapes is of paramount importance in ecology, taxonomy and paleobotanical studies. Classification systems proposed by domain experts support these descriptions. Despite the importance of these visual descriptive systems, classifications based on this expert's knowledge may be ambiguous or limited when representing shapes in unknown scenarios, as expected for biological exploratory domains. This work proposes a novel strategy to automatically discover the shape categories in a set of unlabeled leaves by only using the leaf-shape information. In particular, we overcome the task of discovering shape categories from different plant species for three different biological settings.The proposed method may successfully infer the unknown underlying shape categories with an F-score greater than 92%.The approach also provided high levels of visual interpretability, an essential requirement in the description of biological objects. This method may support morphological analysis of biological objects in exploratory domains.