Classification of Mediterranean crops with multisensor data: per-pixel versus per-object statistics and image segmentation

Abstract

Remote sensing of land cover in Mediterranean regions is complicated by the high landscape diversity, which is typical of both natural and agricultural lands. This spatial complexity reduces the accuracy of the common per-pixel classification of multi-spectral remotely sensed imagery. In this paper we show that per-field statistics derived from multi-spectral imagery enhances separability between different crops and terrain categories. We also found that Synthetic Aperture Radar (SAR) imagery processed with Thematic Mapper (TM) imagery in a synergistic context produces a higher enhancement of discrimination if per-field statistics are used. Finally, we show that image segmentation is a convenient way to apply this approach avoiding field digitizing by computing per-segment statistics of training fields and classifying the segmented image through Linear Canonical Discriminant Analysis.