Natural language understanding remains one of the weak spots of Artificial Intelligence and cognitive systems in general. In cognitive linguistics, image schemas were introduced as spatio-temporal relations learned from sensorimotor processes that constitute conceptual building blocks for highlevel cognition, such as language and reasoning. In this role, they have been successfully employed in computational concept invention and conceptual metaphor research. However, due to their abstract nature identifying them in natural language is an open challenge. To address this issue, this paper proposes a spectral clustering method combined with semantic role labeling to semi-automatically detect image schemas in natural language. In the majority of identified spatial clusters from the Europarl corpus the proposed method detected image schemas, which shows that it works effectively on large corpora. The outcome of this method is a repository of natural language expressions annotated with image schemas that can be used to improve spatial language understanding and provide examples to supervised machine learning approaches.