A Lightweight Spectral-Spatial Unified Feedforward Convolutional Transformer Framework for Spaceborne Hyperspectral Image Classification

Abstract

Large-scale and high-precision classification are key merits of spaceborne hyperspectral technique, providing a vital basis for rapid discovery of diverse anomalies. However, its vast data volume poses major challenges for real-time applications. Compared to classical hyperspectral image classification (HSIC) methods, deep learning (DL) strategies have improved accuracy but still face bottlenecks due to model complexity and cannot be conducted on resource-constrained orbital platforms. Thus, we propose a lightweight framework to achieve the optimal trade-off among memory, time, and accuracy. The key breakthrough adopts a memory-efficient dimensionality reduction (DR) method that reduces the memory cost by over 60% while breaking away from the conventional paradigm that spectral DR inevitably compromise spectral-spatial structures, thus greatly enhancing information fidelity. Moreover, the created convolution-enhanced Transformer (CET) fulfills seamless integration of global–local feature using minimal convolutional blocks on the single encoder, thereby remarkably reducing network complexity and enhancing computational speed. Extensive evaluations demonstrate that our model matches the accuracy of state-of-the-art (SOTA) methods and exhibits remarkably low computational cost (parameters < 35 K, FLOPs < 300 M, GPU memory < 1.5GB), suggesting its potential for onboard deployment. And the devised architecture reduces detection time by at least 20% while sustaining robust capability across varying imaging scenarios. Our results create a new efficiency-accuracy paradigm for real-time HSIC, crucial for achieving instantaneous remote sensing.