Implementation of Deep Learning Algorithm with Residual U-Net Architecture for Building Detection
Ananda Putra, Esti Suryani, Wiharto Wiharto · 2023
Automatic building detection from satellite imagery plays a crucial role in rapidly developing areas. Building detection can assist in illegal building detection, population estimation, and so on. U-Net is a convolutional neural network architecture that is commonly used to detect buildings from satellite imagery. However, the U-Net's deep layer structure makes it vulnerable to vanishing gradients and may affect model performance. To address this issue, residual blocks are applied in the U-Net architecture to overcome the vanishing gradient problem to produce a more accurate model for detecting buildings from satellite imagery. Training and testing on the Massachusetts Building Dataset demonstrate that Residual U-Net achieves an F1 score of 0.842 and an IoU of 0.728, outperforming U-Net with an F1 score of 0.814 and an IoU of 0.688. Moreover, Residual U-Net surpasses other architectures such as FCN, Seg-Unet, and HFSA U-Net in terms of performance.