Abstract: Aiming at the problem that it is difficult to obtain the roof area of buildings accurately and efficiently in the evaluation of roof photovoltaic resources,the FPN＿AttentionUnet semantic segmentation network is proposed to realize high-precision automatic extraction of building roofs. The network integrates soft attention mechanism and double-layer feature pyramid FPN to extract accurate semantic information and refine segmentation results. The Soft-Attention mechanism is used to process and connect the feature map of the encoding part and the decoding part. Double-layer feature pyramid FPN fuses feature maps of different scales to obtain feature information of different scales. The unmanned aerial vehicle is used to obtain the building data set over a certain area of Suzhou and the WHU public data set of Wuhan University for training,respectively. The training results show that compared with Unet,AttentionUnet and FPNUnet networks,the proposed FPN＿AttentionUnet has higher accuracy in building outer contour extraction,which effectively improves the effect of edge extraction. In the self-made dataset,the category pixel accuracy CPAreaches 95.56%,and the average intersection and union ratio MIoUreaches 91.10%. In the WHU public dataset,the segmentation effect is also better than other comparison networks. Finally,taking Changzhou Campus of Hohai University as an example,the proposed algorithm was used to segment buildings from UAV images to evaluate the photovoltaic power generation and photovoltaic module installation potential of the area.