Incorporating contextual information into object-based image analysis workflows

Object-based approaches to image analysis have achieved considerable prominence in the last decade and are now widely considered superior to pixel-based approaches, particularly when extracting features from high-resolution remotely sensed data. The oft-cited advantage of the object-based approach is the ability to simultaneously incorporate spectral, geometric, textural, and contextual information into the classification process. However, context has been ignored in many applications of object-based techniques, despite its importance to human cognition and the current technical capacity to accommodate it. We attribute this oversight to reliance on linear approaches to image analysis and argue that iterative approaches, while more complex, can produce more stable classifications and lead to improved accuracy. We provide examples from four recent land-cover mapping projects that show how context - the relative position of individual objects to neighbor objects - was used to improve feature discrimination in heterogeneous landscapes. We also show how this key factor in pattern recognition was combined with data fusion techniques to maximize object discrimination and to exploit existing investments in remote-sensing data (e.g., imagery, LiDAR, and vector GIS datasets). Although inclusion of contextual information in object-based image analysis presents both analytical and processing challenges, we found that the benefits of improved accuracy and landscape representation far outweigh potential costs.