SDHSI-Net: Learning Better Representations for Hyperspectral Images via Self-Distillation

TL;DR

The paper tackles hyperspectral image classification under limited labeled data and high spectral dimensionality. It proposes SDHSI-Net, a lightweight 3D-2D backbone with a Teacher head and two auxiliary Student heads trained online through self-distillation, plus a triplet loss to enforce discriminative embeddings. Inputs are treated as $\mathbf{I} \in \mathbb{R}^{H\times W\times C}$ and reduced by PCA to $\mathbf{I}_{\text{PCA}} \in \mathbb{R}^{H\times W\times B}$, from which patches of size $S\times S\times B$ are sampled. The learning objective combines $\mathcal{L}_{\text{CE}}^{\text{T}}$, $\mathcal{L}_{\text{CE}}^{\text{S}}$, $\mathcal{L}_{\text{Logit}}$, $\mathcal{L}_{\text{Hint}}$, and $\mathcal{L}_{\text{Trip}}$ as $\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{CE}}^{\text{T}} + \mathcal{L}_{\text{CE}}^{\text{S}} + \lambda_{\text{L}}\mathcal{L}_{\text{Logit}} + \lambda_{\text{H}}\mathcal{L}_{\text{Hint}} + \lambda_{\text{trip}}\mathcal{L}_{\text{Trip}}$, trained with AdamW and cosine annealing. Experiments on Indian Pines and Salinas show competitive Overall Accuracy, Average Accuracy, and Cohen's $\kappa$ relative to SOTA while achieving lower parameter counts and faster inference, validating effective spectral–spatial learning under data scarcity.

Abstract

Hyperspectral image (HSI) classification presents unique challenges due to its high spectral dimensionality and limited labeled data. Traditional deep learning models often suffer from overfitting and high computational costs. Self-distillation (SD), a variant of knowledge distillation where a network learns from its own predictions, has recently emerged as a promising strategy to enhance model performance without requiring external teacher networks. In this work, we explore the application of SD to HSI by treating earlier outputs as soft targets, thereby enforcing consistency between intermediate and final predictions. This process improves intra-class compactness and inter-class separability in the learned feature space. Our approach is validated on two benchmark HSI datasets and demonstrates significant improvements in classification accuracy and robustness, highlighting the effectiveness of SD for spectral-spatial learning. Codes are available at https://github.com/Prachet-Dev-Singh/SDHSI.

Figures (2)

• Figure 1: The architecture diagram of our SDHSI-Net, where $T$, $S_1$, and $S_2$ are the outputs from the Teacher, Student 1 and Student 2 modules, respectively. We denote the fully connected layer with FC. The distillation takes place between the Teacher $T$ and Students ($S_1$ and $S_2$) using the logit distillation and hint distillation losses, described in Section \ref{['sec:method_proposed_model']}. Also, we incorporated the triplet loss for better separability of classes.
• Figure 2: The classification maps obtained from our proposed model SDHSI-Net on a) Indian Pines and b) Salinas datasets. Here, GT, $S_1$, $S_2$ and $T$ denote the ground truth, Student 1, Student 2 and Teacher, respectively.