Standardized object-based dual CNNs for very high-resolution remote sensing image classification and standardization combination effect analysis

Abstract

ABSTRACT Advances in the object-based convolutional neural network (CNN) have demonstrated the superiority of CNNs for image classification. However, any object-based CNN, regardless of its model structure, only stacks the square images with different scales when extracting features. The impact of background information around the segmented object (the number of pixels around the segmented object) for the classification accuracy is neglected. In addition, blurred object boundaries and feature representation, as well as huge computational redundancy, restrict the application for very high-resolution remote sensing image (VHRI) classification. To solve these problems, a novel standardized object-based dual CNN (SOD-CNN) is proposed for VHRI classification. First, based on geographic object-based image analysis, the image is segmented into homogeneous regions. Second, these less-segmented objects are over-segmented into superpixels with high compactness to provide crisp and accurate boundary delineation at the pixel level. Third, four standardization methods are developed to limit the number of pixels around the segmented object. The standardized less-segmented object and over-segmented object are fed into two different CNNs to capture different perspectives of features at local and global scales. Finally, feature fusion based on the full connection is performed to integrate the class-specific classification results. The effectiveness of the proposed method was verified by using two VHRI, which achieved excellent classification accuracy, consistently outperforming the benchmark comparisons. The overall and per-class classification accuracy was investigated under different standardization combinations. We found that (1) the proposed standardization method not only reduced redundancy of information in the object-based CNN but also highlighted the features of segmented objects; (2) different segmented objects had different optimal standardization combinations; and (3) the classification accuracy was reasonably controlled by the foundation number of training samples.