Onboard Real-Time Hyperspectral Subpixel Mapping With Pushbroom Neural Network

Satellite hyperspectral images provide fine spectral information but often suffer from coarse spatial resolution. This may result in mixed pixels containing multiple land cover classes, degrading the performance of conventional classification methods. Subpixel mapping seeks to assign class labels to multiple subpixels within each coarse pixel, producing a super-resolved classification map. Existing approaches rely on priors that are either handcrafted or obtained via large deep learning architectures for offline processing requiring substantial memory and computational resources. In this work, we focus on the processing setting directly onboard the satellite, where such priors are not available and there are strict constraints on computational complexity. In particular, we propose a line-based deep subpixel mapping network tailored for real-time deployment on satellites with a processing strategy that aligns with the pushbroom acquisition of hyperspectral sensors, named Deep Pushbroom SubPixel Mapping (DPSPM). The model predicts a super-resolved classification map exploiting spatial-spectral context from neighboring lines, along with long-range information from previous lines through efficient state-space models. The method achieves fast inference and has low memory requirements compatible with resource-constrained satellite platforms. Experiments demonstrate that the proposed method achieves state-of-the-art subpixel hyperspectral mapping accuracy, while supporting real-time inference at the line acquisition rate, as tested on a low-power accelerator.