Minimizing Manual Labeling Effort for The Semantic Segmentation of Aerial Images

Many modern applications rely on machine learning to fulfill their purpose. However, machine learning, especially the popular deep learning, requires a sufficient amount of labeled data to train models. For some tasks and in some domains, such as aerial images, labeling data is very time-consuming and thus expensive. We therefore propose strategies using unsupervised learning techniques to identify a subset of the input data which actually needs to be labeled by an expert in order to train a well-performing model. With our strategies, which involve less manual labeling effort, we were able to reduce the amount of training data required to 16%. At the same time, the model trained with this small subset achieved better semantic segmentation performance (average accuracy increase: 0.6%, average mIoU increase: 1.3%) for aerial images than a model trained with the full dataset.