BCD-Net: building change detection based on fully scale connected U-Net and subpixel convolution

In recent decades, building change detection (BCD) algorithms have made significant progress by introducing deep learning (DL) for pixel-based change detection (CD). However, they still suffer the problem of low accuracy owing to insufficient feature extraction that is not discriminative enough. Hence, this results in poor semantic segmentation, including irrelevant change information and inconsistent boundaries. To tackle these problems, we present a novel DL-based method known as building change detection (BCD-Net) in an end-to-end manner. Our approach not only deals with the above-mentioned challenges but also significantly increases the level of accuracy. Moreover, our proposed method is inspired by full-scale U-Net3+ that uses an encoder and decoder for semantic segmentation. We modified the U-Net3+ by adding subpixel convolution layers instead of upsampling layers. BCD-Net is applied in three main parts: (1) data preparation, (2) model training/optimization of model parameters, and (3) building change detection based on tuned models. To evaluate the performance of BCD-Net, the benchmark unmanned aerial vehicle (UAV), and satellite imagery were employed. The BCD-Net obtained an accuracy of 98.80% accuracy with the known WHU building dataset and 94.48% on the EGY-BCD dataset. The obtained results demonstrated that the proposed BCD-Net outperformed the rest of the techniques and achieved competitive accuracy, a low rate of miss-detection, and a false alarm rate.