Performance evaluation of maximum likelihood SAR segmentation for multi-temporal rice crop mapping

Optimal (maximum likelihood) processing for a SAR image comprised of discrete regions of constant radar cross section is now well known. This scheme considerably improves upon windowed and iterative schemes by merging regions on an individual pixel basis. In theory, rigorous expressions for ´false alarm rate´ can be defined but they are perhaps too sensitive to the underlying assumptions of independence and homogeneity. Images can be visually improved by a restraint on the ´surface tension´ of the segmented regions but, to avoid subjective judgement, performance is assessed using pixel-accuracy ground truth from a rice growing area in central Japan based on multi-temporal, 8 m resolution Radarsat data. Gaussian expectation maximisation is used to achieve accurate, fully-unsupervised classification of the area without fixed-value thresholding. This large scale backscatter information is then used in a Bayesian merging scheme to give a significant improvement in performance.