An agent-based extension for object-based image analysis for the delineation of irrigated agriculture from remote sensing data

The newly developed image interpretation approach of agent-based image classification combines the advantages of object-based image classification and expert knowledge. Agent-based classification identifies meaningful objects by autonomous software units that alter their spatial extent and composition to adapt to a changing environment and data availability. Agents deliver highly variable classification results of a remotely sensed scene. Although the approach has proven its general ability, the use of agent-based image classification studies is sparse. With this study, we want to introduce this concept to water resource management in form of the detection of irrigated agriculture. In this study, we present the fundament of an agent-based image classification framework in terms of agriculture and irrigation management that shows promising results. In contrast to pixel-based classification approaches, the agent-based classification uses the shape and the relation of an image object to other image objects to improve classification results. To incorporate the possibility of erroneous classification due to threshold behaviour, a strict and a soft formulation for class membership are applied. The results show that the object- and agent-based approaches both deliver similar results as a traditional pixel-wise classification approach, but improve the completeness of the classes. In this case study, the different formulations have only little influence on the general results, but remain a promising addition to the classification approach. The here presented classification strategy is sensitive to changes in the underlying classification scheme. Nevertheless, this framework is an ideal addition to the toolset of image interpretation especially in natural systems that are affected by human behaviour like irrigated agriculture.