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

• DOI: 10.1080/01431161.2020.1742946
• Publication Date: 2020-06-17T12:59:46Z
• Authors: Liangzhi Li; Ling Han; Huijuan Hu; Zhiheng Liu; Hongye Cao
• Keywords: Contextual image classification; Feature (linguistics); Benchmark (surveying)

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