Binary Neural Network for Hyperspectral pansharpening

ICLR 2026 Conference Submission681 Authors

Keywords: Deep Learning, Hyperspectral Pansharpening, Image Fusion

Abstract: Hyperspectral pan-sharpening aims to generate high-resolution hyperspectral (HRHS) images by fusing low-resolution hyperspectral (LRHS) data with high-resolution panchromatic (PAN) images, enabling applications in mapping, surveillance, and environmental monitoring. While deep learning methods achieve strong performance, their heavy computational and memory demands limit deployment on resource-constrained satellite platforms. To address this, we explore binary neural networks (BNNs) for hyperspectral pan-sharpening. Conventional binarization, however, introduces gradient instability and severe information loss, compromising spectral–spatial fidelity. We propose the Adaptive Tan Identity Straight-Through Estimator (ATISTE), a soft binarization strategy that decouples forward approximation from gradient propagation and employs adaptive scaling to preserve consistency with full-precision features. Building on ATISTE, we design HS-BiNet, a lightweight binary CNN with residual connections and multi-scale fusion to effectively capture spectral–spatial dependencies, while avoiding computationally intensive operations such as unfolding inference and non-local self-attention, thereby ensuring suitability for real-time deployment on edge and satellite platforms. Extensive experiments show that HS-BiNet consistently outperforms binary baselines and remains competitive with, and in some cases surpasses, full-precision models, offering a practical solution for high-fidelity HRHS reconstruction.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 681