Buildings, Approximate True Orthophoto Construction From Satellite Imagery Using Semantic Segmentation and the ICP Algorithm

Orthorectification corrects geometric distortions in satellite imagery, ensuring each point on the orthophoto represents true coordinates on Earth accurately, crucial for urban planning and environmental monitoring. The paper introduces a novel building orthorectification method combining semantic segmentation with deep neural networks and the Iterative Closest Point (ICP) Algorithm, to produce approximate true orthophotos of buildings in satellite images. The proposed method was tested on a newly created dataset of satellite RGB images of Maribor, Slovenia. The method employed a pretrained RefineNet model for semantic segmentation, and performed the ICP algorithm for up to 8 iterations to align building rooftops accurately with their official outlines. Benchmarking against ground truth showed significant improvements in accuracy and Intersection Over Union (IOU), with the best results being achieved within 2–4 iterations of the ICP. The highest average IOU achieved was 0.911 with two iterations, and the highest average accuracy was 0.955 with three iterations. A slight performance decrease with an increased number of iterations was noted, attributed to errors and oversimplifications in the official building outlines.