Building panoptic change segmentation with the use of uncertainty estimation in squeeze-and-attention CNN and remote sensing observations

Abstract

ABSTRACT In conventional frameworks, the building change prediction of the deep model can be just blindly assumed accurate, but sometimes the truth is not. In this study, a new deep convolutional neural network (CNN) is proposed for building panoptic change segmentation using uncertainty estimation in squeeze-and-attention CNN and remote sensing observations. The setup is based on a large-scale dataset, called the LEVIR building change detection dataset (also known as LEVIR-CD), including bi-temporal red-green-blue (RGB) images labelled for building change segmentation with a period of 5 to 14 years have significant land-use changes from 20 various areas that sit in Texas of the United States. The quantitative assessments of the LEVIR-CD dataset show that the panoptic quality (PQ), recognition quality (RQ), segmentation quality (SQ), and mean intersection over union (mIoU) for building panoptic change segmentation are about 91.8, 94.7, 96.9, and 97.3, respectively. Compared with the deep learning networks with different backbones and loss functions, the proposed method demonstrates better performance and good generalization ability for building panoptic change segmentation.