Automatic detection and estimation of the area of buildings

This research is an approach to how deep learning networks can be used as a framework to detect different small objects in a satellite image with high pre- cision and estimate the construction area of the segmented buildings. This solution is applied to a problem in the region of Cyprus where the critical concern is the regulation of urban infrastructures and settlements. However, estimating the building area can be a real challenge if such a building has an unusual shape or is too close to another one. The U-Net architecture has been developed for image segmentation and small object recognition, obtaining fast and precise results. Here, we show the improvement in com- bining this convolutional network for house detection and segmentation with manual annotation. The results indicate that the proposed method can make an improvement in segmentation accuracy by about 10% following an annotation methodology. Furthermore, the mean intersection over union (IoU) score of about 91% validates the promising performance of this model.