Snow-induced noise removal in LiDAR measurements via an enhanced sparse sampling transformer

LiDAR measurements in snowy environments are fundamentally challenged by airborne snowflakes, which appear as spatially sparse, irregular, and predominantly high-frequency artifacts, in sharp contrast to the dense and structured background scene. Conventional denoising operators and fixed-pattern learning-based mechanisms are ill-suited for such noise, as their receptive fields are not adaptive to sparse and non-uniform distributions. To address this challenge, we propose a sparsity-aware Transformer framework that focuses on separating sparse snow-induced artifacts from complex background structures through adaptive and semantically guided sampling. The core of the proposed approach is a sparse sampling deformable attention module, which employs learnable sampling offsets to selectively attend to snow-corrupted regions while suppressing responses from dense background points. To stabilize and guide this adaptive sampling process, auxiliary supervision derived from snowflake masks is introduced, injecting explicit sparsity priors into attention learning. Furthermore, motivated by the frequency-domain discrepancy between sparse noise and structured scene content, a discrete wavelet transform-based feature refinement module and a wavelet-domain sparsity loss formulated in a self-supervised manner are incorporated to reinforce feature-level separation between noise and background. Extensive experiments on both synthetic benchmarks and real-world snowy LiDAR datasets demonstrate that the proposed method consistently outperforms existing filter-based and learning-based approaches, achieving more effective noise suppression while preserving structural integrity of the point clouds.