SAR image classification using deep features and neighborhood heterogeneity confidence refined spatial constraints

ABSTRACT

Recently, frameworks fusing the convolutional neural network (CNN) and Markov random field (MRF) models have been successfully used in synthetic aperture radar (SAR) image classification. Over-smoothing of the details and the incapability to eliminate regional-level misclassification are two common drawbacks of these frameworks, which are caused by the use of conventional neighbourhood consistency based spatial constraints. To address these problems, a neighbourhood heterogeneity confidence – Markov random field (NHC-MRF) and CNN-based framework is proposed for SAR image classification. In this framework, an NHC index is constructed in the binary energy function of the NHC-MRF to refine the conventional spatial constraints by adaptively allowing the existence of heterogeneity in the neighbourhood. The NHC index consists of a label coexisting probability term and a top-2 label term, which are constructed based on the outputs of the CNN. By using the label coexisting probability term, the details can be protected by allowing heterogeneity, and the regional-level misclassification can be eliminated by adjusting the impact of the neighbourhoods by using label coexisting probability based weighting. The top-2 label term is used as a correction of the coexisting probability term considering the disturbance of speckle noise. The NHC-MRF is further fused with a CNN by constructing the unary energy term and initial labels based on the outputs of the CNN. The effectiveness and superiority of the proposed framework are experimentally demonstrated using three SAR datasets. The experimental results demonstrate that the superiority of the NHC-MRF is derived from the simultaneous realization of retaining details and eliminating regional-level misclassification.

KEYWORDS:

• Convolutional neural network
• Synthetic aperture radar image classification
• neighborhood heterogeneity confidence
• Markov random field

Acknowledgements

This work was supported by the National Natural Science Foundations of China (grant 62101206), the Department of Education Foundation of Anhui Province (grant KJ2021A0021), the National Key Research and Development Program of China (Grant No.2019YFE0115202), and the Open Research Fund of the National Engineering Research Center for Agro-Ecological Big Data Analysis and Application (Grant No.AE202215). We thank LetPub (www.letpub.com) for its linguistic assistance during the preparation of this manuscript.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Funding

This work was supported by the National Key Research and Development Program of China [2019YFE0115202]; National Natural Science Foundations of China [62101206]; Open Research Fund of National Engineering Research Center for Agro-Ecological Big Data Analysis and Application [AE202215]; Department of Education Foundation of Anhui Province [KJ2021A0021].