Bayesian Markov Chain Random Field Cosimulation for Improving Land Cover Classification Accuracy

Abstract

This study introduces a Bayesian Markov chain random field (MCRF) cosimulation approach for improving land-use/land-cover (LULC) classification accuracy through integrating expert-interpreted data and pre-classified image data. The expert-interpreted data are used as conditioning sample data in cosimulation, and may be interpreted from various sources. The pre-classification can be performed using any convenient conventional method. The approach uses the recently suggested MCRF cosimulation algorithm (Co-MCSS) to take a pre-classified image as auxiliary data while performing cosimulations conditioned on expert-interpreted data. It was tested using a series of expert-interpreted data sets and an image data set pre-classified by the supervised maximum likelihood (SML) algorithm. Results show that with the density of the interpreted data (pixel labels) increasing from 0 to 1.81 % of total pixels, the accuracy of optimal classification maps from Co-MCSS improves by 8.49 to 20.96 %, being much higher than that generated by SML and those purely conditioned on expert-interpreted data. This means that expert-interpreted data may largely contribute to the accuracy of LULC classification from remotely sensed imagery, and inversely the pre-classified image data also can largely contribute to the accuracy of LULC classes simulated by the MCRF approach based on expert-interpreted data. Therefore, the proposed approach can improve classification accuracy of pre-classified maps as long as some expert-interpreted data are available. The main advantage of this approach is that it may comprehensively utilize a variety of available information seamlessly to LULC classification through expert interpretation and spatial cosimulation.