Comparing Active Learning for Satellite Imagery: Contrastive Learning, Diversity, Confidence Sampling, and Cold Start Strategies

In this paper, we investigated and compared common active learning strategies for classification and semantic segmentation tasks using satellite imagery. Our findings suggest that the initialization phase of the iterative active learning process can impact performance. Specifically, we observe that dropout-based methods require careful application, as early-iteration query models may struggle to select informative samples, particularly for classification tasks. Additionally, we introduce a GitHub repository providing a test environment for evaluating active learning methods tailored to satellite imagery. Using both class-balanced and class-imbalanced setups from the EuroSAT classification dataset, as well as a subset of the Dynamic World semantic segmentation dataset, our experiments provide insights to help researchers refine active learning approaches for satellite imagery applications.