Uncertainty-based image segmentation using U-Net architecture and Monte Carlo Dropout
DOI:
https://doi.org/10.34185/1562-9945-4-165-2026-04Keywords:
damaged building detection, convolutional neural networks, computer vision, semantic segmentation, Monte Carlo Dropout, ASPP, U-Net, Google EarthAbstract
Historically, convolutional neural networks have been widely used to solve image object recognition problems, in particular for preliminary damage analysis after natural disasters. The U-Net architecture has become one of the popular solutions for image segmentation; in recent years, many studies and modifications of this model have proposed, using both attention mechanisms and separate modules to improve classification results. At the same time, research into uncertainty assessment and explanation of the operation of segmentation models still need to be deepened, which is why this work is relevant.
Although increasing accuracy and efficiency is an important component of any model, in tasks such as preliminary damage and loss analysis, the assessment of confidence, or uncertainty, is one of the decisive factors, since it can be used for further decision-making. The purpose of the study is to build a segmentation model of the modified U-Net architecture with the ASPP module, taking into account uncertainty by multiple runs of the model with different dropout masks (Monter Carlo Dropout) for preliminary analysis of damaged buildings in images.
The paper presents a study of the modified U-Net architecture with the replacement of the bottleneck with the ASPP module and the use of Monte Carlo Dropout. Based on the calculation formulas of variance, entropy and mutual information, uncertainty maps were formed, which were used to improve segmentation. It is shown that ASPP increases the accuracy of object shape detection, and the combined loss function contributes to improving segmentation. In the experiments, an increase of 3.8% of the average IoU was achieved, 4% for surviving and 5.3% for destroyed buildings. The uncertainty masks were analyzed and used in the method with two threshold values of confidence and model uncertainty to select the most reliable areas. For the practical implementation of the described methods and approaches, there is a need for their further study.
References
Moreno González, C. L., Montoya, G. A., & Lozano Garzón, C. (2025). Toward reliable post-disaster assessment: Advancing building damage detection using You Only Look Once convolutional neural network and satellite imagery. Mathematics, 13, 1041. https://doi.org/10.3390/math13071041.
Liu, J., Luo, Y., Chen, S., Wu, J., & Wang, Y. (2024). BDHE-Net: A novel building dam-age heterogeneity enhancement network for accurate and efficient post-earthquake assessment using aerial and remote sensing data. Applied Sciences, 14, 3964. https://doi.org/10.3390/app14103964.
Mazurets, O., Molchanova, M., Shurypa, M., & Sobko, O. (2026). Devising a neural-network method for assessing the condition of destroyed buildings using images from un-manned aerial vehicles. Eastern-European Journal of Enterprise Technologies, 1(2(139)), 6–16. https://doi.org/10.15587/1729-4061.2026.351605
Deng, L., & Wang, Y. (2022). Post-disaster building damage assessment based on im-proved U-Net. Scientific Reports, 12, 15862. https://doi.org/10.1038/s41598-022-20114-w.
Yang, Q., Wang, Z., Liu, S., & Li, Z. (2024). Research on improved U-net based remote sensing image segmentation algorithm. arXiv. https://doi.org/10.48550/arXiv.2408.12672.
He, K., Zhang, X., Ren, S., & Sun, J. (2015). Spatial pyramid pooling in deep convolu-tional networks for visual recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 37(9), 1904–1916. https://doi.org/10.48550/arXiv.1406.4729.
Gal, Y., & Ghahramani, Z. (2016). Dropout as a Bayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33rd International Conference on Machine Learning (Vol. 48, pp. 1050–1059). https://doi.org/10.48550/arXiv.1506.02142.
Rey, M., Mnih, A., Neumann, M., Overlan, M., & Purves, D. (2025). Uncertainty evalua-tion of segmentation models for Earth observation. arXiv. https://doi.org/10.48550/arXiv.2510.19586.
Dechesne, C., Lassalle, P., & Lefèvre, S. (2021). Bayesian U-Net: Estimating uncertainty in semantic segmentation of Earth observation images. Remote Sensing, 13(19), 3836. https://doi.org/10.3390/rs13193836.
Saidu, C., & Csató, L. (2021). Active learning with Bayesian U-Net for efficient semantic image segmentation. Journal of Imaging, 7(2), 37. https://doi.org/10.3390/jimaging7020037.
Saifullah, S., & Dreżewski, R. (2026). Bayesian optimization-driven U-Net architecture tuning for brain tumor segmentation. In Proceedings of the 6th International Electronic Con-ference on Applied Sciences (Eng. Proc., 124(1), 22). https://doi.org/10.3390/engproc2026124022.
Downloads
Published
Issue
Section
License
Copyright (c) 2026 System technologies
This work is licensed under a Creative Commons Attribution 4.0 International License.
