A Performance Analysis of U-Net and U-Net++ in Building Footprint Extraction Using XAI

Abstract

In recent years, it has been reported that deep neural networks produce higher accuracies in the extraction of building footprints compared to traditional methods. However, this achievement is mainly attributed to their capacity to discern and model intricate spatial patterns presented through remotely sensed data. The opaque nature of the decision-making processes within deep learning models hinders the interpretability of their outputs, thereby casting doubt on their reliability. To address this challenge, U-Net and U-Net++ models were used to extract building footprints from the dataset created from SPOT imagery, and Saliency, Gradient SHAP and Integrated Gradients explainable artificial intelligence (XAI) algorithms were employed to make the decision-making mechanisms of the methods understandable and explainable. The Intersection over Union (IoU) value for the U-Net++ model was calculated as 0.9646, while it was 0.9598 for the U-Net model. In summary, the U-Net++ model outperformed the U-Net model in terms of the metrics considered. Additionally, according to the XAI results, it was found that the U-Net++ model focused more on specific regions while the U-Net model showed a more random distribution. As a result, it is seen that the U-Net++ model has a more distinct decision-making mechanism. © 2025 Elsevier B.V., All rights reserved.