Automatic rice-crop mapping using maximum likelihood SAR segmentation and Gaussian expectation maximisation

Accurate, large scale crop monitoring requires the use of weather-independent sensors such as SAR. Rice is the staple food in many parts of Asia and knowledge of rice growth provides valuable economic and environmental information. Extensive pixel accuracy ground truth from an area of Kojima, Japan is used to compare the accuracy of unsupervised mapping algorithms. An acceptable classification is achieved by Gaussian expectation maximisation applied to existing maximum-likelihood (ML) based segmentation methods. Using a Bayesian extension of the ML segmentation scheme gives a dramatic improvement in accuracy which results in a pixel classification accuracy of 86% (68% kappa) and an essentially exact estimate of rice coverage within a 1000 hectare area.