Abstract: Geographic object-based image analysis (GEOBIA) techniques have recently seen considerable development in comparison to traditional pixel-based image analysis, representing a paradigm shift in remote sensing interpretation. The main aim is to incorporate and develop geographic-based intelligence. The random forest (RF) machine learning method is a relatively new, non-parametric, data-driven classification method that can supply intelligent means for feature selection and classification modelling. This paper presents a novel RF GEOBIA method for land-cover classification that makes full use of the advantages of GEOBIA and RF. A detailed RF GEOBIA workflow is proposed to guide the design and implementation of the method, and to guide experts during elaboration of feature selection and classification modelling. Theoretical and experimental results are compared with the support vector machine (SVM) approach, demonstrating that it is a robust and intelligent method for land-cover classification with wrapper feature selection and classification modelling. The RF GEOBIA method reduces the number of features required, computing time, and memory requirements, with no associated reduction in performance. It also provides a priori knowledge for further classification and supports large scale applications where "big data" is involved.