SpecAL: Towards Active Learning for Semantic Segmentation of Hyperspectral Imagery

We investigate active learning towards applied hyperspectral image analysis for semantic segmentation. Active learning stems from initially training on a limited data budget and then gradually querying for additional sets of labeled examples to enrich the overall data distribution and help neural networks increase their task performance. This approach works in favor of the remote sensing tasks, including hyperspectral imagery analysis, where labeling can be intensive and time-consuming as the sensor angle, configured parameters, and atmospheric conditions fluctuate. In this paper, we tackle active learning for semantic segmentation using the AeroRIT dataset on three fronts - data utilization, neural network design, and formulation of the cost function (also known as acquisition factor, uncertainty estimator). Specifically, we extend the batch ensembles method to semantic segmentation for creating efficient network ensembles to estimate the network’s uncertainty as the acquisition factor for querying new sets of images. Our approach reduces the data labeling requirement and achieves competitive performance on the AeroRIT dataset by using only 30% of the entire training data.