Relating the land-cover composition of mixed pixels to artificial neural network classification output

Artificial neural networks are attractive for use in the classification of land cover from remotely sensed data. In common with other classification approaches, artificial neural networks are used typically to derive a ''hard'' classification, with each case (e.g., pixel) allocated to a single class. However, this may not always be appropriate, especially if mixed pixels are abundant in the data set. This paper investigates the potential to derive information on the land-cover composition of mixed pixels from an artificial neural network classification. The approach was based on relating the activation level of artificial neural network output units, which indicate the strength of class membership, to land-cover composition. Two case studies are discussed which illustrate that the activation level of the artificial neural network outputs themselves were not strongly related to pixel composition. However, re-scaling the activation levels, to remove the bias towards very high and low strengths of class membership imposed by the unit activation function, produced measures that were strongly related to the land-cover composition of mixed pixels. In both case studies, significant correlations (all r > 0.8) between the re-scaled activation level of an output unit and the percentage cover of the class associated with the unit were obtained.