Monitoring Urban Changes in Mariupol/Ukraine in 2022/23. (arXiv:2309.08607v1 [cs.CY])

The ability to constantly monitor urban changes is of large socio-economic
interest. Previous works have already shown approaches in this field with the
use of Deep Neural Networks (DNNs) and transfer learning. However, they fell
short in demonstrating temporal scale outside of either the training or
transfer domain.

This work builds on existing research and proves that transfer learning with
the use of historic data is a feasible solution, which still allows the urban
change monitoring of later years. We considered a case with limited access to
public and free Very High Resolution (VHR) imagery to guide the transfer. To
provide a high temporal resolution, the core data of our monitoring method
comprised multi-modal Synthetic Aperture Radar (SAR) and optical multispectral
observations from Sentinel 1 and Sentinel 2, respectively.

We chose a practical application of our methods for monitoring urban-related
changes in the city of Mariupol in Ukraine during the beginning of the
Russo-Ukrainian War in 2022/23. During this conflict, availability of VHR data
was limited and hence an inexpensive direct transfer to the years 2022/23 was
rendered impossible. Instead, a transfer was made for the years 2017-2020 that
provided sufficient public and free VHR data with an application of the
transferred model in the years late 2021 to mid-2023. It was shown that
transferring for the years 2017-2020 with this inexpensive historical VHR data
enabled monitoring during times of war in 2022/23.

An ablation study on the impact of the frequency of observations showed our
method as resilient to even a large loss of observations. However, it also
indicated that our method, despite the multi-modal input, was more dependent on
optical observations than SAR observations. Neither the indirect transfer, nor
the malfunction of Sentinel 1B had a significant impact on the monitoring
capabilities of our method.