Knowledge Based Image Analysis of Agricultural Fields in Remotely Sensed Images

The application problem that is addressed in this contribution is to monitor large agricultural areas for landuse, crop types and crop yield forecasting, using remotely sensed data. Visual image interpretation is too slow and does not produce quantitative results fast enough and bottom up image analysis is unlikely to succeed in terms of the complexity of multitemporal, multisource data. We investigate knowledge based, hypothesis driven image analysis. Hypotheses about surface objects are generated from prior geometric and radiometric model knowledge such as should be available from a 4-dimensional geo information system 4dGIS. Hypotheses are evaluated against a selection of the data and model parameters are estimated. We use a predictor – corrector method of hypothesis and parameter estimation with the expected cost of misclassification of surface elements as the optimization criterion. The predictor model generates (abstracted) RS data from a 3dim+time scene model, a sensor model, an illumination and an atmosphere model. Equivalences between the general modelling and parameter estimation method and existing image processing and pattern recognition algorithms are explored. Feature extraction is shown to be equivalent to the inverse modelling of image data to radiometric and geometric model parameters for some cases of spatial feature extraction. We report on mapping geometric, radiometric and process knowledge onto physical models, how to acquire and use prior knowledge, the role of sensor models and on an experiment on the extraction of agricultural fields from airborne synthetic aperture radar data of Flevoland in The Netherlands.