Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

SpectralTrain: A Universal Framework for Hyperspectral Image Classification

Meihua Zhou, Liping Yu, Jiawei Cai, Wai Kin Fung, Ruiguo Hu, Jiarui Zhao, Wenzhuo Liu, Nan Wan

TL;DR

SpectralTrain addresses the efficiency bottleneck in hyperspectral image classification by introducing a spectral curriculum that starts with PCA-based spectral downsampling and progressively restores full spectral information. The framework is architecture-agnostic and couples a compute-budgeted training schedule with CPU-based spectral compression, achieving 2–7× speedups with minimal accuracy loss across Indian Pines, Salinas-A, and CloudPatch-7. Key contributions include a universal training paradigm compatible with CNNs, 3D spectral networks, and transformers, a formal cost- and stage-switching model, and extensive experiments demonstrating robustness across spatial scales, spectral characteristics, and climate-related cloud classification. The work highlights training strategy optimization as a powerful complement to architectural design in hyperspectral learning, with code released at the provided GitHub repository.

Abstract

Hyperspectral image (HSI) classification typically involves large-scale data and computationally intensive training, which limits the practical deployment of deep learning models in real-world remote sensing tasks. This study introduces SpectralTrain, a universal, architecture-agnostic training framework that enhances learning efficiency by integrating curriculum learning (CL) with principal component analysis (PCA)-based spectral downsampling. By gradually introducing spectral complexity while preserving essential information, SpectralTrain enables efficient learning of spectral -- spatial patterns at significantly reduced computational costs. The framework is independent of specific architectures, optimizers, or loss functions and is compatible with both classical and state-of-the-art (SOTA) models. Extensive experiments on three benchmark datasets -- Indian Pines, Salinas-A, and the newly introduced CloudPatch-7 -- demonstrate strong generalization across spatial scales, spectral characteristics, and application domains. The results indicate consistent reductions in training time by 2-7x speedups with small-to-moderate accuracy deltas depending on backbone. Its application to cloud classification further reveals potential in climate-related remote sensing, emphasizing training strategy optimization as an effective complement to architectural design in HSI models. Code is available at https://github.com/mh-zhou/SpectralTrain.

SpectralTrain: A Universal Framework for Hyperspectral Image Classification

TL;DR

SpectralTrain addresses the efficiency bottleneck in hyperspectral image classification by introducing a spectral curriculum that starts with PCA-based spectral downsampling and progressively restores full spectral information. The framework is architecture-agnostic and couples a compute-budgeted training schedule with CPU-based spectral compression, achieving 2–7× speedups with minimal accuracy loss across Indian Pines, Salinas-A, and CloudPatch-7. Key contributions include a universal training paradigm compatible with CNNs, 3D spectral networks, and transformers, a formal cost- and stage-switching model, and extensive experiments demonstrating robustness across spatial scales, spectral characteristics, and climate-related cloud classification. The work highlights training strategy optimization as a powerful complement to architectural design in hyperspectral learning, with code released at the provided GitHub repository.

Abstract

Hyperspectral image (HSI) classification typically involves large-scale data and computationally intensive training, which limits the practical deployment of deep learning models in real-world remote sensing tasks. This study introduces SpectralTrain, a universal, architecture-agnostic training framework that enhances learning efficiency by integrating curriculum learning (CL) with principal component analysis (PCA)-based spectral downsampling. By gradually introducing spectral complexity while preserving essential information, SpectralTrain enables efficient learning of spectral -- spatial patterns at significantly reduced computational costs. The framework is independent of specific architectures, optimizers, or loss functions and is compatible with both classical and state-of-the-art (SOTA) models. Extensive experiments on three benchmark datasets -- Indian Pines, Salinas-A, and the newly introduced CloudPatch-7 -- demonstrate strong generalization across spatial scales, spectral characteristics, and application domains. The results indicate consistent reductions in training time by 2-7x speedups with small-to-moderate accuracy deltas depending on backbone. Its application to cloud classification further reveals potential in climate-related remote sensing, emphasizing training strategy optimization as an effective complement to architectural design in HSI models. Code is available at https://github.com/mh-zhou/SpectralTrain.

Paper Structure

This paper contains 54 sections, 5 theorems, 31 equations, 8 figures, 8 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. Training Strategy in Computer Vision
  4. Curriculum Learning
  5. Hyperspectral Imaging for Climate and Environment
  6. Method
  7. Computational Constraints and Optimization Objectives
  8. Training Schedule and Stage Switching
  9. Constraint Factors and Training Process
  10. Resolution-controlled steps.
  11. Spectral compression prior to upload.
  12. Stage switching.
  13. Computational Constraint Order Search
  14. Results
  15. Experimental Setup
  16. ...and 39 more sections

Key Result

Proposition 1

Let $\mathcal{T}_{H}:f\mapsto f*h$. Since $|H|\le 1$, $\|\,\mathcal{T}_{H}f\,\|_{L^{2}}\le\|f\|_{L^{2}}$ and $\|\mathcal{T}_{H}f-\mathcal{T}_{H}g\|_{L^{2}}\le\|f-g\|_{L^{2}}$. Thus pairwise separations in $L^{2}$ contract band-wise under LFC.

Figures (8)

  • Figure 1: Overview of SpectralTrain. (a) Discrete decision-making based on RGB sample difficulty (e.g., selecting easier low-frequency content first). (b) Hyperspectral imaging-based curriculum (ours): a continuous transformation $T_t(\cdot)$ progressively introduces complex spectral and spatial patterns via PCA-based spectral reduction and image-level downsampling, gradually increasing data complexity as training proceeds.
  • Figure 2: Motivation example for a spectral curriculum on Indian Pines. (A) Average spectra of two clusters show band-localized discriminative cues, so uniform band dropping risks removing signals. (B) PCA cumulative explained variance illustrates that a small number of components preserve most energy, motivating an information-preserving low-cost start; PCA is used here as a representative linear compressor. (C) Per-epoch transfer and compute scale with the number of retained components $k$, so beginning with small $k$ saves early cost. (D) Curriculum stages progressively increase both spectral components and spatial size (PC1/PC2/PC3 composites). Panels illustrate one instantiation with PCA; Table \ref{['tab:dr']} shows that alternative reducers (UMAP/ICA) produce comparable behavior under the same staged schedule.
  • Figure 3: OA/AA across backbone scales under SpectralTrain.
  • Figure 4: Kappa across ResNet and ConvNeXt variants under SpectralTrain.
  • Figure 5: Supplementary Figure S1. Class distribution in CloudPatch-7. Bars show per-class sample counts (c01–c07), revealing moderate imbalance.
  • ...and 3 more figures

Theorems & Definitions (5)

  • Proposition 1: Contraction under LFC
  • Proposition 2: Irrecoverable high-frequency energy loss
  • Theorem 3: Eckart--Young--Mirsky
  • Proposition 4: Nyquist-consistent sampling per PC
  • Proposition 5: Separation versus approximation