Simple Synthetic Data as Source Domain for Transfer Learning to Remote Sensing as a Target Domain

Züfle, Andreas2023-06-142023-06-14https://hdl.handle.net/1920/13313Deep Learning continues to grow as a prevalent toolset among multiple disciplines, including Remote Sensing and image analysis. Correspondingly, to more easily apply the deep neural networks to different subject matter domains, Transfer Learning, from natural image datasets, including ImageNet, has become a de-facto method for many Deep Learning applications, including Remote Sensing. However, such an approach may have limitations related to the differences on the characteristics of natural photographic image datasets and the characteristics of Remote Sensing. This study aims to determine if a fairly arbitrary, easily produced set of synthetic datasets can be iteratively developed and used for Transfer Learning for a typical Deep Learning task. We found this is readily and surprisingly feasible.masters thesesenCopyright 2022 Brian L. ShawSimple Synthetic Data as Source Domain for Transfer Learning to Remote Sensing as a Target DomainTransfer learningRemote sensingSynthetic dataDeep neural networkExplainable artificial intelligenceConvolutional neural network