MAF-UNet: Building Extraction Algorithm Based on Improved U-Net Network

Extracting building information from remote sensing images holds significant reference value for advancing smart city construction. However, the challenges associated with complex backgrounds, diverse building types, and interference from buildings of different scales contribute to a decrease in the building extraction accuracy. A multi-scale aggregation fusion U-Net (MAF-UNet) network was proposed based on the U-Net architecture. By incorporating a constructed Frequency-andCoordinate attention module to suppress irrelevant information and concatenate it with low-level features in the decoder, the network addressed the challenges posed by the diverse background. Within the bridging module, the integration of the Receptive Field Block module facilitated the extraction of multiscale features across an expanded receptive field. Additionally, an Adaptive Spatial Feature Fusion module was incorporated into the decoder, enabling the adaptive learning of weights for features at varying scales. Finally, the Rectified Linear Unit activation function was replaced by the Gaussian Error Linear Unit activation function, aiming to enhance the convergence speed and overall performance during the training process. Experimental validation demonstrated that the constructed MAF-UNet network outperformed the U-Net network on the WHU building dataset, achieving improvements of
$\mathbf{0 . 4 8 \%}$
in accuracy,
$4.8 \%$
in precision,
$1.86 \%$
in F1-score, and
$3.19 \%$
in intersection over union. On the INRIA aerial image labeling dataset, the constructed MAF-UNet network achieved improvements of
$\mathbf{0 . 3 4 \%}$
in accuracy,
$2.98 \%$
in precision,
$\mathbf{1 . 5 2 \%}$
in F1-score, and
$\mathbf{2 . 3 0 \%}$
in intersection over union than these of the U-Net network.