| Abstract [eng] |
Change detection is one of many topics in the remote sensing domain. Advancements in this topic aid various use cases spanning from infrastructure planning or urbanization tracking to environmental problem-solving and monitoring. Generally, remote sensing has gained a lot of attraction from researchers. Various models that have scored impressive results in the computer vision domain have also shown noteworthy accomplishments in the remote sensing area. However, popularity distribution seems to decay when coming to more concrete applications such as global warming change detection. Both, global and national attraction is a little scarce and one of the main problems could be challenges that come in the remote sensing domain: optical aberrations, data quality, seasonal changes interference and knowledge transfer difficulties. With all things in mind, this study has proposed medium analysis which aims to filter out noisy data and modified an existing neural network. The model was build to employ knowledge sharing by incorporating Prithvi's model feature maps into the detection pipeline. Its performance was slightly better across all 3 test cases: original OSCD dataset, MBSC dataset and manually gathered imagery of Lithuania. The objectives of this work include analysing relevant literature and datasets, investigating and proposing methods for working with multi-spectral images, developing new neural network by using existing algorithms and evaluating proposed solutions, providing recommendations. The hypothesis being tested is that it is possible to successfully track changes using existing satellite imagery resources. |
