Classification of remotely sensed images is often based on assigning classes on a 'per-pixel' basis and therefore ignoring spatial information captured by the image. In this paper a method is proposed that combines spatial and spectral information in two steps. Spatial information is first extracted by segmentation: spectral homogeneous regions in the image are identified on the basis of similarity of the entire spectral shape from visible to shortwave infrared. These homogeneous areas are classified and next, this information is used to classify the remaining heterogeneous pixels. The method is applied and validated on DAIS images acquired over an area covered with natural vegetation and agricultural activities in southern France. Results indicate significant improvements of classification results compared to traditional per-pixel classifiers.