Training Methods of Multi-Label Prediction Classifiers for Hyperspectral Remote Sensing Images

Abstract

Hyperspectral remote sensing images, with their amalgamation of spectral richness and geometric precision, encapsulate intricate, non-linear information that poses significant challenges to traditional machine learning methodologies. Deep learning techniques, recognised for their superior representation learning capabilities, exhibit enhanced proficiency in managing such intricate data. In this study, we introduce a novel approach in hyperspectral image analysis focusing on multi-label, patch-level classification, as opposed to applications in the literature concentrating predominantly on single-label, pixel-level classification for hyperspectral remote sensing images. The proposed model comprises a two-component deep learning network and employs patches of hyperspectral remote sensing scenes with reduced spatial dimensions yet with a complete spectral depth derived from the original scene. Additionally, this work explores three distinct training schemes for our network: Iterative, Joint, and Cascade. Empirical evidence suggests the Joint approach as the optimal strategy, but it requires an extensive search to ascertain the optimal weight combination of the loss constituents. The Iterative scheme facilitates feature sharing between the network components from the early phases of training and demonstrates superior performance with complex, multi-labelled data. Subsequent analysis reveals that models with varying architectures, when trained on patches derived and annotated per our proposed single-label sampling procedure, exhibit commendable performance.