Lightweight Masked Spectral Spatial Self Supervised Learning for Few Shot Hyperspectral Image Classification
Year: 2026
Type: preprint
Abstract: Hyperspectral image (HSI) classification benefits from hundreds of contiguous spectral bands, yet reliable pixel-level annotations remain costly and scarce. Self-supervised learning (SSL) is able to decrease label dependency, although the most recent models of HSI are based on computationally intensive Transformers, graph-based, multi-branch, or contrastive networks, and might not actively optimize spatial differentiation. This paper proposes Masked SpectralSpatial Self-Supervised Learning (MSS-SSL), which is a relatively simple framework that simultaneously handles spectral recovery, spatial discrimination, few-shot label efficiency, and computational efficiency. MSS-SSL collaboratively undergoes masked spectral-band reconstruction, spatial patch contrastive representation learning and lightweight spectral-spatial encoding. A 128-dimensional representation is generated using a depthwise-separable spectral stem and three depthwise-separable spatial residual blocks that do not use a large Transformer backbone. In pretraining, spectral groups that are adjacent to each other are masked and solved using unlabeled patches, with two spatially semantics-invariant views matched to an InfoNCE loss. The trained encoder is then refined using 1, 3, 5, and 10 pixels that are labeled by the classes. The experimental routine includes Indian Pines, Pavia University, Salinas, and Houston 2013 and tests the total, mean, and Cohen kappa, macro-F1, trainable, FLOPs, inference latency, throughput and label efficiency. Matched masked and contrastive and joint-objective ablations are worked with identical encoder capacity, so that the effects of objectives of objective nature can be disentangled with architectural capacity. The number of repeated seeds is five, to provide statistical stability. This resulting design has a controlled basis of testing the hypothesis of whether complementary reconstruction and discrimination goals may act to exploit the accuracy-to-computation ratio when label abundance becomes extreme; numerical performance assertions would only be made intentionally when the experiments were finished.
Source: Research Square
Institution: Xidian University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7235911758
