Active Learning-Based Spectral-Spatial Classification for Discriminating Tree Species in Hyperspectral Images

Exploiting spectral-spatial information and reducing the number of required training samples are important for improving tree species classification performance in hyperspectral images. In this article, an active learning-based spectral-spatial classification (ALSSC) model is proposed to reduce the demand for training samples while improving the classification performance. To improve classification performance, the proposed ALSSC employs two ways to exploit spectral-spatial information within the hyperspectral image: 1) features used in classification are extracted from multiscale superpixels; 2) the classification result is refined by guided filtering and subsequently employed as the input for the next round of classification. To reduce the demand for training samples, after each round of classification, active learning (AL) is adopted to select the most informative samples from the unlabeled testing set to enrich the training set. To validate the effectiveness of the proposed ALSSC, experiments are conducted using a tree species classification dataset collected by an airborne hyperspectral sensor. Remarkably, when compared to the state-of-the-art AL-based approach using the same number of labeled samples, the ALSSC demonstrates an accuracy improvement of 11.62%. In addition, trained with fewer labeled samples, the ALSSC outperforms state-of-the-art spectral-spatial classification methods that do not incorporate AL.