Active Learning with Deep Support Vector Data Description for Earth Observation Satellite Image Classification

This paper presents a novel Active Learning (AL) framework that integrates deep Support Vector Data Description (SVDD) for optical satellite image classification. By defining a decision boundary within the deep SVDD model, we implement and iteratively update AL strategies based on this boundary to select the most informative samples for labeling. Our approach significantly reduces the need for labeled data while maintaining high classification accuracy. We evaluate our framework on the EuroSAT dataset using both RGB and Multispectral (MS) bands, demonstrating that our AL strategies namely, Most Similar, Top-K, and Bottom-K can effectively approximate the performance of a model trained on the full dataset with a fraction of the samples. The Top-K and Bottom-K strategies, in particular, show remarkable efficiency in querying informative samples. This study advances the application of AL in RS, offering a cost-effective solution for high-dimensional Earth Observation data analysis.