PurposeThe wind loading on a building is likely to deviate further from the known wind loading due to the complexity of the real-world land coverage. To address this issue, research is needed in two separate areas. First, wind tunnel testing needs to be conducted for more complex terrains. Second, research is needed to classify real-world land coverage with high accuracy, specifically for wind engineering applications. This paper deals with this second area of research. The machine learning-based land cover prediction is a promising technique because it can remove subjectivity in human interpretation of upwind terrain.Design/methodology/approachThis paper presents a new deep neural network for land coverage prediction that can distinguish low- and mid-rise buildings in the built environment to enhance the estimation of surface roughness necessary in wind engineering. For the dataset, Landsat 8 satellite images were used. A patch-based convolutional neural network was employed and improved. The network predicted the land coverage at the center of the patch. Two different label schemes were used where the proposed network either achieved better accuracy than the conventional model or recognized additional building types while maintaining a similar level of accuracy.FindingsCompared to the validation accuracy of 78% in a previous study, the proposed method achieved the validation accuracy of 90% thanks to the improvements made in this study as well as the consolidation of labels with similar surface roughness. When additional building categories were added, the validation decreased to 80%, which is comparable to the previous study but is now able to predict different building types.Originality/valueThe improvement of the proposed method will depend on the site characteristics. For the sites tested in this paper, the error reduction in wind speed and pressure was up to about 55%. In addition to more accurate wind speed and pressure, better identification of buildings will benefit wind engineering research, as different building types cause different downwind effects. An example application would be automated recognition of areas that have a certain distance from the target building type to identify downwind areas affected by high winds.
Read full abstract