Adaptive conditional random field classification framework based on spatial homogeneity for high-resolution remote sensing imagery

Since the conditional random field (CRF) model can integrate spectral and spatial-contextual information of high spatial resolution (HSR) remote sensing images in a unified framework, it becomes an effective approach to optimize the classification results. However, the results of traditional classification methods based on the CRF are sensitive to the parameters. In this paper, an adaptive conditional random field (ACRF) model is designed to utilize the spatial information more flexibly and improve the accuracy. In the ACRF, the spatial homogeneity is employed to achieve adaptive parameters control, which can evaluate the effect of the unary potentials and pairwise potentials of different pixels. Two datasets are used in the experiments, and the results demonstrate that the proposed method can improve the classification accuracy, alleviate salt-and-pepper noises, and retain detailed information. Compared with other methods, ACRF shows a better performance for HSR image classification, integrating the spatial-contextual and spectral information.