An innovative deep learning-based approach for evaluating the urban rooftop solar photovoltaic potential

Accurate access to urban-scale rooftop information is crucial for assessing rooftop solar photovoltaic(PV) potential. Deep learning semantic segmentation methods, which combine convolutional neural networks(CNNs) with vision transformers(ViTs), have demonstrated significant promise in extracting building rooftops from satellite images. However, these methods still face challenges in terms of limited feature extraction capabilities, high computational complexity, and labor-intensive data labeling. To overcome these limitations, this study proposes a network called TFAViT. This network employs a dual-branch fusion architecture that processes CNN and ViT in parallel. It replaces the standard convolutions in the CNN branch with frequency-adaptive dilated convolution(FADC) and alterable kernel convolution(AKC) to enhance feature representation. TFAViT adopts top-k sparse attention(TKSA) to reduce the computational complexity. Furthermore, this study proposes a method for optimizing the sampling region based on the specific characteristics of the target region, which enhances model performance without increasing the number of labeled samples. The methodology was validated by an application in Baohe District, Hefei, China. The results show that the IoU and F1 scores of the TFAViT model have increased by 1.31% and 0.80%, respectively. The method of optimizing sampling region resulted in notable improvements in the model's performance, with an increase of 2.09% in the F1 score and 3.36% in the IoU. This method successfully extracted a total rooftop area of 22.96 km2 in Baohe District of Hefei City, with a corresponding rooftop solar photovoltaic potential of 3843.99 GWh annually.