IMPROVEMENT OF U-NET MODEL SEGMENTATION WITH ACCOUNT OF UNCERTAINTY
DOI:
https://doi.org/10.34185/1991-7848.itmm.2026.01.071Keywords:
damaged building recognition, convolutional neural networks, computer vision, semantic segmentation, Monte Carlo Dropout, ASPP, U-NetAbstract
Despite the importance of increasing accuracy and efficiency, uncertainty assessment, which is used for decision-making, plays a key role in the tasks of preliminary damage analysis. The aim of the study is to improve the segmentation of the modified U-Net architecture with the ASPP module and taking into account uncertainty through Monte Carlo Dropout (MCD) method.
The paper proposes replacing the bottleneck with ASPP and applying stochastic dropout. Based on variance, entropy and mutual information, uncertainty maps are formed that improve segmentation. It is shown that ASPP increases the accuracy of determining object shapes, and the combined loss function improves the quality of the results. In the experiments, an increase of 3.8% of the average IoU, 4% for intact and 5.3% for destroyed buildings was achieved. The positive impact of the proposed methods and approaches on the segmentation capabilities of the model is confirmed, while further research is needed for practical implementation.
References
Yang Q., Wang Z., Liu S., Li Z. Research on improved U-net based remote sensing image segmentation algorithm. arXiv, 2024. URL: https://doi.org/10.48550/arXiv.2408.12672 (data zvernennia: 25.03.2026).
He K., Zhang X., Ren S., Sun J. Spatial pyramid pooling in deep convolutional networks for visual recognition // IEEE Transactions on Pattern Analysis and Machine Intelligence. 2015. Vol. 37, No. 9. P. 1904–1916. URL: https://doi.org/10.48550/arXiv.1406.4729 (data zvernennia: 25.03.2026).
Gal Y., Ghahramani Z. Dropout as a Bayesian approximation: Representing model uncertainty in deep learning // Proceedings of the 33rd International Conference on Machine Learning. 2016. Vol. 48. P. 1050–1059. URL: https://doi.org/10.48550/arXiv.1506.02142 (data zvernennia: 26.03.2026).
Rey M., Mnih A., Neumann M., Overlan M., Purves D. Uncertainty evaluation of segmentation models for Earth observation. arXiv, 2025. URL: https://doi.org/10.48550/arXiv.2510.19586 (data zvernennia: 26.03.2026).
