Multiscale Discriminative Feature Adaptive Learning Network for Hyperspectral Image Classification

Deep learning has advanced hyperspectral image (HSI) classification by efficiently extracting spectral and spatial features. However, its performance is often limited when labeled data are scarce and class distributions are imbalanced. Moreover, current methods frequently fail to effectively capture multi-scale discriminative representations that bridge local spatial details with global contextual information. To address these challenges, in this work we present a Multiscale Discriminative Feature Adaptive Learning (MDFAL) network for HSI classification. MDFAL employs a cascaded architecture that integrates shallow, intermediate, and deep sub-networks, jointly optimized through multi-scale feature outputs and adaptive weighting of multi-task losses. To capture discriminative representations more effectively, the framework incorporates three complementary modules: (i) a spectral–spatial attention module (SSAM) that strengthens cross-dimensional interactions while suppresses redundancy; (ii) a ConvLSTM-based contextual learning module (CCLM) that models both short- and long-range spectral dependencies together with spatial context; and (iii) a dilated convolution block that expands the receptive field, thereby enhancing multi-scale representation across network layers. Experimental results demonstrate that the proposed MDFAL model surpasses existing methods by effectively capturing multi-scale discriminative features with yielding strong intra-class compactness and inter-class separability. Under identical conditions, MDFAL improves overall accuracy by 2.31%, 3.12%, and 3.10% on the Indian Pines, Pavia University, and Tea datasets, respectively. These improvements arise from the synergistic interaction between its cascaded shallow, intermediate, deep architecture and adaptive multi-scale loss weighting. The findings confirm MDFAL as a robust framework for advancing HSI classification network design.