Active Multiple Kernel Fredholm Learning for Hyperspectral Images Classification

Active learning (AL) represents an encouraging solution for hyperspectral image classification based on domain adaptation (DA) with very limited labeled samples in target domain. Although the traditional AL methods have exhibited the promising results in DA, some challenges still exist. On the one hand, the previous AL schemes assign a label to the most informative unlabeled data by user and, thus, are characterized by errors, time, and costs, which ignores dealing with noisy and complex data in target domain. On the other hand, the traditional AL methods based on kernel prediction model assume a predefined kernel and the identical distribution for source and target domains, which reduces the performance of classifier on target domain. To overcome these issues, we propose the Active Multiple Kernel Fredholm Learning (AMKFL), where a Fredholm kernel regularized model is presented to label the samples instead of the user, and then define two Fredholm integrals with multiple kernels to find an optimal kernel between different distributions, which increases the classification accuracy and generalization capabilities in noisy cases. The experimental results with two popular hyperspectral data sets show that the proposed AMKFL improves the classification accuracy significantly compared to the traditional methods while decreasing the user interaction.

Automated summary

Active learning is an effective solution for hyperspectral image classification with very limited labeled samples, but traditional methods face challenges. To overcome these issues, the study proposes Active Multiple Kernel Fredholm Learning, which uses a Fredholm kernel regularized model and multiple kernels to improve classification accuracy.