Classification of multitemporal remote sensing data of different resolution using Conditional Random Fields

The increasing availability of multitemporal optical remote sensing data offers new potentials for land cover analysis. We present a novel approach for enhancing the classification accuracy of medium resolution data by combining them with high resolution data of an earlier acquisition time, thus saving data acquisition costs. Our approach uses Conditional Random Fields to model both spatial and temporal dependencies. Temporal context is considered by a novel extension of the CRF concept by an additional temporal interaction potential, which can model dependencies between identical regions in images of different acquisition times and scales. The model also considers different levels of abstraction in the class structures at different scales. The approach is tested with two set-ups of Ikonos, RapidEye, and Landsat imagery.