Digital Image-Based Building Detection Using Ghost Encoder-Decoder

Automatic building detection is a challenge in the field of Geoinformatics due to buildings that blend in with their surroundings. To address this issue, a neural network method called Ghost Encoder-Decoder, based on the U-Net framework, has been proposed. The method utilizes residual blocks and ghost bottlenecks, a convolutional arrangement that reduces the number of parameters. This results in more optimal performance and efficient memory utilization. The Ghost Encoder-Decoder was trained and tested using high-resolution satellite images, achieving an F1-Score of 85.45% and an Intersection-over-Union (IoU) of 74.63%. These results outperform other methods such as PSPNet, DeepLabv3+, RefineNet, SegNet, Bayesian-SegNet, and BRRNet. Additionally, the Ghost Encoder-Decoder has a comparatively lower parameter count of 13.76 million, making it a more lightweight and cost-effective model.