WeatherMamba: Reliable Geometry-Aware Generalization for Adverse-Weather LiDAR Sensor Point-Cloud Semantic Segmentation Authors: He Huang, Xintai Zhang, Yidan Zhang, Junxing Yang Published: 2026 ABSTRACT Adverse weather degrades LiDAR semantic segmentation by altering return density, measured intensity, local geometry, and feature reliability. We study source-only domain generalization from clear-weather or synthetic training data to unseen real adverse-weather scenes. WeatherMamba is a geometry- and reliability-aware state-space framework that integrates multiscale neighborhood aggregation, reliability-conditioned feature refinement, and efficient long-range sequence modeling. Its central component, Weather-Guided Reliability Gating (WGRG), derives a continuous point-wise gate from conditionally normalized measured intensity, multiscale local density, and latent features. This gate bounds residual corrections without an externally supplied weather type, prior target-reflectance values, or deployment-time meteorological measurements. A controlled pseudo-weather strategy models signal perturbation and point-support loss separately during source-domain training. WeatherMamba achieves 37.4% mIoU on SemanticKITTI→SemanticSTF and 22.8% on SynLiDAR→SemanticSTF, ranking third overall while attaining the highest class IoU in five classes under each protocol. In the single-seed module ablation, the complete configuration improves the module-free baseline from 29.4% to 37.4%. Relative to PTv3 with Modules, WeatherMamba reduces parameters, FLOPs, latency, and peak allocated memory by 19.1%, 34.8%, 16.5%, and 25.2%, respectively, with a 0.4-point reduction in mIoU. These results indicate a competitive accuracy–efficiency trade-off under the evaluated protocols. 1. Introduction LiDAR point-cloud semantic segmentation assigns a semantic label to each 3D point and supports localization, mapping, obstacle reasoning, and motion planning in autonomous-driving systems. Although current networks perform well on clear-weather benchmarks, their reliability can deteriorate in rain, fog, and snow. Controlled sensor studies show that adverse weather changes the number of valid returns, measured intensity, target visibility, and detection range. Increasing rain can reduce both detected-point counts and measured return intensity [1,2]. Fog and rain also reduce automotive LiDAR target-detection performance [3,4], while controlled snow experiments report attenuation, backscattering, false detections, and sensor-surface contamination [5]. These effects alter the geometric and radiometric evidence available to a segmentation network, creating a distribution shift between source-domain training data and adverse-weather deployment data. Figure 1 illustrates representative SemanticSTF [6] point clouds under fog, rain, and snow conditions. We consider domain generalization rather than target-domain adaptation. The model is trained on clear-weather or synthetic source-domain data and evaluated directly on unseen real adverse-weather scenes. No SemanticSTF target-domain frame is used for model optimization. This setting reflects the practical difficulty of collecting and labeling every combination of sensor, road scene, precipitation level, visibility, and target material. A deployed autonomous system must, nevertheless, remain reliable when weather conditions differ from those represented during training. Adverse-weather generalization presents several coupled difficulties. Attenuation and point loss weaken local geometric continuity, especially around thin structures, object boundaries, and distant targets. Sparse artifacts and unstable returns can produce unreliable local features that are then amplified by deeper layers. When local evidence is incomplete, recognition also depends more strongly on long-range context. Degradation also varies continuously and cannot be represented adequately by a simple light/heavy label. A practical model should, therefore, infer degradation cues from the observed point cloud while preserving local geometry and long-range context. Recent methods address different aspects of adverse-weather LiDAR domain generalization. Weather-Aware Autopilot introduces adaptive normalization, dual-attention fusion, and proxy-label learning for SemanticKITTI/SynLiDAR→SemanticSTF transfer [7]. Park et al. analyze geometric perturbation and point drop and design targeted augmentation for adverse-weather segmentation [8]. No Thing, Nothing addresses the disproportionate degradation of safety-critical “things” classes through superclass binding and beam-wise feature distillation [9]. PhyDAWS uses physically inspired weather simulation for domain-generalized point-cloud segmentation [10]. Together, these studies support feature adaptation, targeted augmentation, safety-critical class modeling, and physically inspired simulation. Observation-derived continuous point-wise degradation, explicit local geometric support, and efficient long-range context, however, have not been studied jointly. We address this gap with WeatherMamba, a geometry- and reliability-aware state-space framework for adverse-weather LiDAR segmentation. Its central component, WGRG, represents degradation as an observation-derived continuous point-wise gate, not as a discrete weather-class decision or a physically calibrated severity estimate. WGRG combines measured intensity, multiscale local density, and latent-feature evidence to bound residual corrections using only the observed scan and source-trained quantities at inference. MANF aggregates multiscale neighborhoods to preserve local geometric support, RADM refines features according to point-wise reliability without removing points, and a hierarchical selective state-space backbone propagates long-range context efficiently. During training, weather-motivated augmentation exposes the source model to controlled signal perturbation and support-loss patterns. Each component, therefore, addresses a distinct failure mode within the same source-only framework. We evaluate two source-only transfer protocols using SemanticKITTI or SynLiDAR as the source and SemanticSTF as the common unseen target. The experiments report aggregate, class-wise, and weather-wise accuracy; complete module combinations; WGRG evidence sensitivity; pseudo-weather strength sensitivity; and a matched PTv3 [11] efficiency comparison. This evaluation examines both segmentation accuracy and the computational cost of observation-only deployment. Our main contributions are summarized as follows:WeatherMamba integrates MANF, RADM, and a hierarchical Mamba backbone to address local geometric support, feature reliability, and long-range context within a unified source-only architecture.WGRG derives a continuous point-wise gate from measured LiDAR intensity, multiscale density, and latent-feature evidence. It bounds residual corrections without an external weather label, prior target-reflectance values, or deployment-time meteorological input.A weather-motivated augmentation strategy controls signal perturbation and point-support loss separately. Pre-specified default strengths are evaluated through a post hoc sensitivity analysis over neighboring settings.Under two SemanticSTF transfer protocols, the complete configuration improves the module-free baseline by 8.0 points. Relative to PTv3 with Modules, it reduces parameters, FLOPs, latency, and peak memory with a 0.4-point mIoU trade-off. WeatherMamba integrates MANF, RADM, and a hierarchical Mamba backbone to address local geometric support, feature reliability, and long-range context within a unified source-only architecture. WeatherMamba integrates MANF, RADM, and a hierarchical Mamba backbone to address local geometric support, feature reliability, and long-range context within a unified source-only architecture. WGRG derives a continuous point-wise gate from measured LiDAR intensity, multiscale density, and latent-feature evidence. It bounds residual corrections without an external weather label, prior target-reflectance values, or deployment-time meteorological input. WGRG derives a continuous point-wise gate from measured LiDAR intensity, multiscale density, and latent-feature evidence. It bounds residual corrections without an external weather label, prior target-reflectance values, or deployment-time meteorological input. A weather-motivated augmentation strategy controls signal perturbation and point-support loss separately. Pre-specified default strengths are evaluated through a post hoc sensitivity analysis over neighboring settings. A weather-motivated augmentation strategy controls signal perturbation and point-support loss separately. Pre-specified default strengths are evaluated through a post hoc sensitivity analysis over neighboring settings. Under two SemanticSTF transfer protocols, the complete configuration improves the module-free baseline by 8.0 points. Relative to PTv3 with Modules, it reduces parameters, FLOPs, latency, and peak memory with a 0.4-point mIoU trade-off. Under two SemanticSTF transfer protocols, the complete configuration improves the module-free baseline by 8.0 points. Relative to PTv3 with Modules, it reduces parameters, FLOPs, latency, and peak memory with a 0.4-point mIoU trade-off. 2. Related Work 2.1. LiDAR Semantic Segmentation and Long-Range Context Modeling LiDAR semantic segmentation methods are commonly organized into point-based, projection-based, voxel-based, and global-context architectures. PointNet++ [12] and RandLA-Net [13] learn hierarchical local features directly from point sets. Projection-based methods, including SqueezeSegV3 [14] and SalsaNext [15], convert point clouds into range-view representations for efficient 2D processing. Voxel-based methods such as Cylinder3D [16], PolarNet [17], and sparse-convolution networks regularize 3D space and achieve strong outdoor segmentation accuracy. Projection and voxelization, however, can compress or quantize fine geometry, while repeated neighborhood aggregation can be costly for long outdoor scans. Global context becomes especially important when adverse weather removes or corrupts local evidence. Transformer-style self-attention captures long-range dependencies but scales quadratically with sequence length. Structured state-space models offer a more efficient alternative. S4 [18] and Mamba [19] support efficient long-sequence modeling, while Vision Mamba [20] and VMamba [21] adapt selective state-space propagation to visual data. In point clouds, however, serialization can separate physically adjacent points. WeatherMamba, therefore, combines state-space propagation with explicit multiscale neighborhood recovery. 2.2. Physical Effects of Adverse Weather on LiDAR Measurements Sensor-level studies document the local geometry degradation, unstable returns, and measured-intensity shifts introduced by adverse weather. Filgueira et al. [1] reported that increasing rain can reduce LiDAR intensity and the number of detected points. Goodin et al. [2] modeled rain-rate-dependent degradation for automotive LiDAR, and Kutila et al. [3] measured substantial performance loss in controlled fog and rain. Montalban et al. [4] compared automotive LiDAR point clouds in artificial rain and fog, showing that degradation depends on both sensing technology and environmental conditions. Controlled snowfall experiments have also reported attenuation, backscattering, false detections, and sensor contamination [5]. These findings indicate that degradation varies continuously with atmospheric severity, range, sensor characteristics, and scene properties. Throughout this paper, the per-point radiometric value provided by SemanticKITTI and SynLiDAR is termed measured LiDAR intensity (or remission), rather than intrinsic material reflectance. It depends on target properties, range, incidence angle, sensor calibration, and atmospheric propagation. WeatherMamba, therefore, requires no database of target-reflectance values. Measured intensity is used only as one observation-derived cue, together with point density and latent-feature statistics. 2.3. Adverse-Weather Domain Generalization and Data Augmentation Early robustness strategies include geometric filtering and generic data augmentation. Radius/statistical outlier filtering and DROR-style approaches can suppress isolated weather returns but depend on hand-tuned thresholds and may also remove valid sparse structures [22,23]. PolarMix [24] improves LiDAR training diversity by mixing point-cloud sectors, while PointDR [6] regularizes source-domain representations for adverse-condition generalization. UniMix [25] and recent domain-generalization approaches further improve transferability through mixing, uncertainty, or representation learning. More recent work explicitly targets adverse-weather domain generalization. Weather-Aware Autopilot [7] combines adaptive feature normalization, dual-attention fusion, and proxy-label generation, using SemanticKITTI and SynLiDAR as source domains and SemanticSTF as the unseen target. Rethinking Data Augmentation [8] identifies geometric perturbation and point drop as major data-level factors and proposes selective jittering and learnable point drop. No Thing, Nothing [9] shows that safety-critical “things” classes degrade disproportionately and introduces feature binding and beam-wise distillation. PhyDAWS [10] uses physically inspired weather simulation, including weather-specific attenuation and density/occlusion mechanisms, to improve domain-generalized point-cloud segmentation. These methods form the recent comparison context for WeatherMamba. WeatherMamba complements these approaches by placing observation-derived point-wise gating within a geometry- and reliability-aware state-space architecture. Its pseudo-weather branch models sensor-relevant degradation directions but is not intended as an exact atmospheric simulator. The sensitivity analysis consequently reports the strongest settings only within the tested range. 2.4. Methodological Positioning WeatherMamba coordinates long-range sequence modeling with dedicated local-geometry and reliability pathways and an observation-derived gate that requires no discrete weather metadata. The hierarchical Mamba backbone models long-range context, MANF aggregates multiscale local support, RADM performs reliability-conditioned refinement, and WGRG applies bounded point-wise recalibration using measured-intensity, density, and latent-feature evidence. The module and sensitivity studies evaluate these design choices under a fixed source-only protocol. 3. Materials and Methods 3.1. Overall Framework WeatherMamba performs semantic segmentation of LiDAR point clouds affected by weather-induced degradation. As shown in Figure 2, an input batch is represented as P∈RB×N×Din, where B is the batch size, N is the number of sampled points per scan, and Din is the per-point input dimension. Each point contains 3D coordinates (x,y,z), a measured LiDAR intensity/remission value, and optional auxiliary attributes. Measured intensity is treated as a sensor-return attribute, not as intrinsic, pre-known target reflectance. The pipeline has four stages. Feature preparation first maps coordinates, measured intensity, and auxiliary attributes into a latent space. MANF aggregates multiscale neighborhoods to recover local geometric continuity, and RADM estimates point-wise reliability from local statistics to suppress unstable responses softly. The hierarchical WeatherMamba backbone then propagates the refined features through bidirectional selective state-space blocks to capture long-range context. WGRG derives a continuous degradation level for each point from measured-intensity, multiscale point-density, and latent-feature evidence. This level gates a bounded point-wise residual correction without a target-weather class label. Finally, the local and global representations are fused and mapped to point-wise semantic predictions. 3.2. Multiscale Adaptive Neighborhood Fusion Point-cloud serialization enables efficient sequence modeling, but spatial neighbors may be separated in the resulting sequence. Under adverse weather, sparse returns, scattering noise, and partial occlusion further weaken local geometric continuity. The Multiscale Adaptive Neighborhood Fusion (MANF) module precedes the WeatherMamba backbone to recover local structure and provide geometric evidence. As illustrated in Figure 3, MANF receives point features x∈RB×N×C and coordinates p∈RB×N×3. For each point, the feature branch constructs three neighborhoods with ks=8, km=16 and kl=32. The edge feature at each scale is concatenated with the center feature and projected by a scale-specific MLP:(1)Fs=ϕs([x,es]),Fm=ϕm([x,em]),Fl=ϕl([x,el]), where Fs, Fm and Fl denote small-, medium- and large-scale neighborhood features. The three branches represent fine boundaries, intermediate structures, and the more stable geometry obtained from broader neighborhoods. The neighborhood sizes {8,16,32} form a compact geometric progression that covers these scales while keeping neighborhood construction tractable. They remain fixed across all reported experiments and SemanticSTF weather subsets. Here, C is the latent channel dimension of the current local block and is distinct from the number of semantic classes. The multiscale branch explicitly models both aggregation and contrast across receptive fields. The additive multiscale context is defined as(2)Fsum=Fs+Fm+Fl, while the cross-scale difference term is computed as(3)Fdiff=(Fm+Fl)−Fs. The difference term emphasizes patterns that persist across larger neighborhoods while suppressing isolated small-neighborhood responses associated with weather-induced noise. In parallel, the geometry statistics branch computes local mean, local standard deviation and density from the point coordinates. These statistics are projected by a geometry encoder to obtain Ggeo∈RB×N×C. The final fusion input is(4)Fcat=x⊕Fsum⊕Fdiff⊕Ggeo, where ⊕ denotes channel concatenation. The output is generated by a fusion mapping followed by residual normalization:(5)Fout=LayerNormx+ψ(Fcat), where ψ(·) consists of a linear projection, GELU activation, dropout, and a final linear layer. MANF, thus, restores local geometric context before global propagation by the sequence model. 3.3. Reliability-Aware Denoising Module The Reliability-Aware Denoising Module (RADM) suppresses outlier responses in feature space. In contrast to hard-threshold preprocessing, RADM retains all points and estimates their reliability from local geometric statistics. It comprises local-statistics extraction, noise detection and feature refinement, and confidence-guided adaptive fusion. Local geometric statistics extraction: For each point, RADM computes the local mean, standard deviation, point density, and feature-similarity variance within its k-nearest neighborhood. Object surfaces generally exhibit greater local continuity than sparse weather-induced returns. These statistics are concatenated with the original point features to provide structural cues for reliability estimation. Noise detection and feature refinement: An MLP with sigmoid activation processes the original feature Fin and the local statistics to estimate a point-wise noise probability S∈[0,1]N×1. A larger Si indicates lower reliability. A parallel branch uses the local mean context to generate the candidate feature Frepair=Fin+MLP(Fin,Mean), where the residual connection preserves the original semantic information and improves training stability. Confidence-guided adaptive fusion: The noise probability first interpolates between the original and refined features. ConfidenceNet then predicts a point-wise gate G ∈[0,1]N×1 from Fin and S, controlling the contribution of the refined branch:(6)Finter=(1−S)⊙Fin+S⊙Frepair,Fout=G⊙Finter+(1−G)⊙Fin. where ⊙ denotes element-wise multiplication. The two-stage weighting suppresses unreliable responses while retaining sparse but informative points, without requiring manual noise annotations. 3.4. Hierarchical WeatherMamba Backbone The hierarchical WeatherMamba backbone models long-range dependencies with efficient computation. Selective state-space modeling has linear sequence complexity, in contrast to full self-attention, and is, therefore, suitable for large outdoor LiDAR scans. In the stage-wise design shown in Figure 2, each stage contains an input projection, two bidirectional Mamba blocks, depth-wise convolutional local mixing, and LayerNorm. Transition MLPs connect adjacent stages, and a final normalization layer produces the high-level representation processed by WGRG. The unordered point cloud is first serialized according to its spatial coordinates. Each bidirectional Mamba block contains forward and backward scans. The forward branch processes the original sequence, while the backward branch processes its reversal to provide complementary context. Their output is(7)Xl=ϕConcatMambaf(Xl−1),Rev(Mambab(Rev(Xl−1))), where Xl is the feature of the l-th stage, Mambaf(·) and Mambab(·) denote the forward and backward state-space branches, Rev(·) denotes sequence reversal, and ϕ(·) denotes linear projection and nonlinear fusion. State-space scanning propagates long-range information efficiently, but serialization can weaken local physical neighborhoods. Each stage, therefore, applies depth-wise convolutional mixing to restore short-range interactions in the latent sequence. The residual update is(8)X˜l=LNXl−1+DWConv(Xl), where DWConv(·) denotes local depth-wise convolutional mixing and LN(·) denotes LayerNorm. Stacking three stages integrates local geometric detail with long-range semantic context while retaining efficient sequence modeling. 3.5. Weather-Guided Reliability Gating Adverse-weather effects vary across points and are not represented reliably by a weather label assumed to be available at deployment. The Weather-Guided Reliability Gating (WGRG) module shown in Figure 4 instead models degradation as a continuous point-wise residual gate. Given point features F={fi}i=1N and fi∈RC, WGRG estimates di∈[0,1] from measured LiDAR intensity, multiscale local density, and latent features. The resulting value bounds a residual correction. Intensity denotes sensor-provided intensity/remission rather than pre-known intrinsic material reflectance, and the scan-wise mean of di is retained only for diagnostics. A complete atmospheric model would require deployment-time measurements such as rain rate or visibility, as well as sensor-specific calibration and atmospheric parameters unavailable in the standard SemanticSTF inference setting. WGRG, therefore, estimates degradation from the observed point cloud while retaining physically motivated intensity and density cues. This design accommodates sensor-, range-, and scene-dependent effects without assuming meteorological inputs or a fixed mapping from weather intensity to point reliability. Because intensity and density also vary with range, elevation, and sensor sampling, WGRG normalizes them using source-training medians m and median absolute deviations a conditioned on a 5×5 range/elevation grid. These statistics are frozen after fitting. Let b(i) denote the cell containing point i, and let Di,k denote its sensor-origin voxel occupancy at K=2 scales (0.2 and 0.4 m). The normalized evidence is(9)ziI=clipasinh(Ii)−mb(i)Imax(ab(i)I,ε),−8,8,zi,kD=cliplog(1+Di,k)−mb(i),kDmax(ab(i),kD,ε),−8,8, where ε=10−5. Lightweight LayerNorm–GELU heads map the signed deviations and their magnitudes, together with the normalized latent feature, to component levels:(10)ℓiI=σ(hI([ziI,|ziI|])),ℓiD=σ(hD([ziD,|ziD|])),ℓiF=σ(hF(LN(fi))). The feature component is learned without reference to a weather prototype. Noisy-OR combines the three components without a discrete weather decision:(11)di=1−(1−ℓiI)(1−ℓiD)(1−ℓiF). Inactive components are set to zero in the ablation, making di equal to the component level when only one component is active. The normalized feature and active physical evidence predict the correction direction, while di controls its magnitude:(12)ui= [LN(fi),ziI,|ziI|,ziD,|ziD|],ci=tanh(MLPres(ui)),fi′=fi+αdici,α=0.2. Inactive intensity/density terms in ui are set to zero. Since ci∈[−1,1]C, each channel correction is bounded by 0.2di. The final residual layer is initialized to zero, so WGRG begins as an identity mapping. WGRG, therefore, applies a point-wise scalar degradation gate to a point-wise channel correction, rather than using multiplicative channel attention. The two occupancy scales represent fine and broader local support at modest computational cost. The 5×5 range/elevation partition balances condition specificity with the number of source samples available in each cell. The residual factor α=0.2 prevents WGRG from overwriting source-domain features, while the EMA momentum of 0.999 stabilizes the learned feature-evidence targets. Both settings remain fixed across the transfer protocols. During training, corrupted and clean reference branches are generated from the same source scan and aligned by invariant point indices. For matched points, the targets are(13)tiI=clip(|Iic−Iir|/128,0,1),tiD=K−1∑kclip1−Di,kc/(Di,kr+10−6),0,1,tiF=clip(1−cos(fT,ic,fT,ir))/2,0,1, where fT is a stop-gradient exponential-moving-average feature teacher with momentum 0.999. For the active component set A, ti(A)=1−∏j∈A(1−tij). Let V contain the Nv valid matches. The calibration, component, and clean-reference losses are(14)Lcal=1Nv∑i∈VSmoothL1(di,ti(A)),Lcomponent=13Nv∑i∈V∑j∈ASmoothL1(ℓij,tij),Lclean=1Nr∑idir. When WGRG is enabled, its supervision objective is(15)LWGRG=0.5Lcal+0.5Lcomponent+0.0Lrank+0.05Lclean. Here, Lcal aligns the fused degradation score with the target formed from the active intensity, density, and feature evidence. Lcomponent supervises the evidence-specific predictions, and Lclean suppresses degradation scores on clean reference points. Lrank is an optional ranking objective. Its weight is zero in the reported experiments, so it does not affect optimization. The equal 0.5 coefficients balance fused-degradation and component-level supervision, while the smaller 0.05 coefficient regularizes clean-reference scores without dominating the segmentation objective. At inference, WGRG uses only the observed features, coordinates, measured intensity, multiscale density, frozen source-training statistics, and learned parameters. The paired reference, EMA teacher, supervision targets, and weather labels are used only during training. In Section 4.6.1, “intensity-only” and “density-only” refer to the evidence used to form di; the normalized latent feature remains common context for the correction direction. The combined intensity, density, and feature configuration gives the strongest result in the evidence study. 3.6. Weather-Motivated Pseudo-Weather Augmentation Controlled pseudo-weather corruption is applied only to source-domain training scans to reduce the gap between clear/synthetic training data and unseen adverse-weather scenes. The augmentation follows sensor-relevant degradation directions supported by LiDAR studies and recent adverse-weather augmentation methods [1,3,4,8,10]: measured-intensity attenuation and perturbation, range-direction jitter, and loss of valid point support. It is not intended to reconstruct the complete atmospheric process. Separate sensitivity configurations use the field alpha with values in the range of {0,0.5,1.0,1.5}. We denote this discrete configuration value by s below; it is not supplied as a continuous run-time variable. For a point with coordinate xi and measured intensity Ii, the support-preserving branch first applies linear intensity attenuation,(16)Iiatt=(1−a(s))Ii,a(s)=sa0,a0∈{0.15,0.30,0.45}. Impulse noise is then applied to a randomly selected subset of points. Let Mi=⊮[UiI