Active Learning for Hyperspectral Image Classification: A comparative review

Abstract

Machine learning algorithms have demonstrated impressive results for land cover mapping from hyperspectral data. To enhance generalization capabilities of statistical models, active learning (AL) methods guide the annotation of the training data set by querying the most informative samples. The training of the classifier can then be performed on an optimal training data set. We bring under the same framework uncertainty, representativeness, and performance-based AL techniques; conduct a benchmark on state-of-the-art methods and release a toolbox ( <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/Romain3Ch216/AL4EO</uri> ) to allow experimentation with these approaches. The experiments are conducted on various data sets: a toy data set, classic hyperspectral benchmark data sets, and a complex hyperspectral scene. We evaluate the methods with usual accuracy metrics as well as complementary metrics, which allow us to provide guidelines when choosing a relevant AL strategy in a real use case.