To improve the all-weather automatic detection of waterlogging at construction sites, a VGG-SegNet semantic segmentation model based on the SegNet network was proposed. The downsampling network was improved by using the VGG19 convolutional neural network, and the importance-aware loss function (IAL) was adopted as the cross-entropy loss function to achieve pixel-level segmentation of waterlogging. The model’s overall intersection over union (OIoU) reaches 98%, and the F₁ score reaches 96%, both significantly better than the SegNet model. A target pixel size method was proposed for calculating the surface area of waterlogged areas. It construed a perspective projection mapping relationship and uses pixel position and area weights. Construction site tests validated the proposed method, showing an average pixel segmentation accuracy of 97.47% for waterlogged areas and an average absolute error rate of 4.24% for area calculation. The VGG-SegNet model performs well in recognizing waterlogging images at construction sites. Therefore, it has great prospects for promotion and application in the accurate, rapid and intelligent detection of waterlogging at construction sites.