Towards Creating Exotic Remote Sensing Datasets using Image Generating AI

Over the past few years, neural networks have been used more often to solve long lasting challenges. Remote sensing and data classification were some of the fields that have widely depended on this continuously developing technology. In this context, remote sensing data related to places with harsh conditions have been missing, especially the ones related to SAR imagery. Such conditions include deserts, glaciers, and icebergs, where lots of people have lost their lives in, due to the lack of efficient methods of searching and finding these people in such critical timing. Training AI models on similar scenarios to fasten the process can be beneficial, but the lack of data is an obstacle in the way of development such models. In this paper, we propose using image generating AI systems to generate remote sensing datasets that are difficult to collect using normal imagery, thus creating more efficient image classification systems that can be used in scenarios such as locating missing people. Several AI models are discussed in this paper: Dall-E 2, Stable Diffusion and Midjourney, where they are found to vary a lot in terms of the generated images, that could be because of the architecture of the model, and the data they trained on. The overall performance of the AI models is promising. Dall-E 2 performed the best in our tests, followed by Stable Diffusion, and finally Midjourney. This research could open the door to using such models in generating lots of datasets, which might solve crucial problems.