Source-linked AI summary
WaveOp-LiteFM: Lightweight Neural-Operator Flow Matching for Satellite-to-Radar Precipitation Retrieval
Chunlei Shi, Yecheng Zhang, Yufeng Zhu, Dan Niu, Yichao Dong, Yongchao Feng, Junming Hou
TL;DR
Satellite-to-radar retrieval must balance fine sparse-precipitation detail against the cost of repeated high-resolution generative inference. WaveOp-LiteFM keeps conditional flow matching in pixel space but replaces the heavy U-Net velocity backbone with gated spectral-local-wavelet operators, achieving strong benchmark retrieval and scalable large-area inference over China. Its supported scope includes tiled deployment rather than distribution-free generalization.
Problem
Satellite-to-radar retrieval needs to preserve sparse, localized precipitation structures while reducing the computational burden of high-resolution pixel-space generative modeling.
Method
WaveOp-LiteFM uses pixel-space conditional flow matching with spectral, local, and wavelet operator branches plus adaptive branch and skip gating.
Results
WaveOp-LiteFM achieves strong retrieval performance on SEVIR and Southeast China benchmarks and demonstrates large-area inference over China, including Typhoon Bavi.
Takeaways & Limitations
The framework combines direct pixel-space precipitation supervision with a lightweight velocity parameterization for benchmark retrieval and deployment-oriented large-area inference.
Takeaways & Limitations
Tiled inference is a deployment mechanism rather than a claim of distribution-free generalization across changed regions, seasons, sensor geometries, or precipitation regimes.
Abstract
from arXiv · showhide
Satellite-to-radar (S2R) retrieval refers to estimating ground-based radar precipitation from geostationary satellite observations, enabling precipitation monitoring in regions with limited radar coverage. While recent generative flow matching models have greatly advanced retrieval quality, they face a critical trade-off: pixel-space formulations suffer from the prohibitive computational costs of attention-based U-Net velocity networks, whereas latent-space modeling often sacrifices fine precipitation details or struggles with sparse targets. To address this dilemma, we propose WaveOp-LiteFM, a lightweight neural operator flow matching framework for S2R retrieval. Our approach introduces a novel velocity backbone built upon the spectral-local-wavelet (SLW) block, enabling efficient and stable flow matching in pixel space. Specifically, the SLW block disentangles precipitation features into three distinct frequency regimes: (i) the spectral branch captures large-scale stratiform organization; (ii) the local branch models short-range interactions; and (iii) the wavelet branch enhances sharp structures while suppressing noisy high-frequency responses. Building on this design, an input-adaptive gating mechanism dynamically fuses features from the three functional branches. Furthermore, a skip gate efficiently reintegrates encoder features through additive fusion within the decoder, avoiding the costly channel concatenation used in conventional U-Net architectures. Experiments on the SEVIR and Southeast China datasets show that WaveOp-LiteFM achieves state-of-the-art retrieval performance while substantially reducing computational costs. Beyond benchmark evaluation, large-area inference over China, including a recent Typhoon Bavi case, demonstrates that WaveOp-LiteFM maintains reliable retrieval quality in large-scale real-world scenarios.
I. INTRODUCTION
Satellite-to-radar retrieval extends precipitation monitoring beyond incomplete radar coverage, but models must preserve sparse convective detail while remaining efficient on large regional grids. WaveOp-LiteFM addresses this trade-off with lightweight spectral-local-wavelet operators and gated fusion in pixel-space flow matching.
- Motivation: Satellite observations provide frequent wide-area information for radar-like precipitation monitoring where ground radar coverage is limited.Operational challenges include terrain blockage, range effects, heterogeneous quality control, and gaps.
- Proposed approach: WaveOp-LiteFM uses a three-branch spectral-local-wavelet block to model broad organization, local convection, and sharp radar-like details.The branches are combined using adaptive gating, while gated additive skip fusion avoids expensive concatenation.
- Motivation: Pixel-space generative retrieval preserves precipitation detail but conventional U-Net velocity networks impose high computational costs, while latent modeling may smooth sparse cores.The paper frames efficiency and fine-scale precipitation preservation as competing requirements.
- Proposed approach: WaveOp-LiteFM is presented as a lightweight pixel-space flow-matching framework that aims to retain high-fidelity retrieval while improving computational efficiency.Its design targets the compute bottleneck of pixel-space generative retrieval.
- Evaluation: Experiments cover SEVIR and Southeast China benchmarks, with additional tiled inference over China including a recent Typhoon Bavi case.The large-area evaluation is intended to demonstrate scalability and robustness in real-world scenarios.
III. METHOD
The method formulates satellite-to-radar retrieval as conditional flow matching from noise toward radar-like precipitation in pixel space. A compact operator encoder-decoder predicts the conditional velocity while combining global, local, and boundary-sensitive features.
- Problem Formulation: The task maps satellite observations and noisy precipitation states to radar-like precipitation fields under efficiency and sparse-structure preservation requirements.Small displacement or oversmoothing can strongly affect threshold skill, while sampling may be repeated over large grids.
- Framework Overview: WaveOp-LiteFM retains pixel-space conditional flow matching for direct precipitation supervision and replaces the heavy U-Net velocity backbone with a lightweight operator-style encoder-decoder.The network receives noisy state y_t, flow time t, and satellite condition c, then predicts the velocity field.
- Framework Overview: Spectral-local-wavelet blocks combine operators for storm-scale organization, local convective interaction, and boundary-sensitive precipitation details.This decomposition provides complementary representations within the velocity network.
- Pixel-Space Conditional Flow Matching: The conditional flow-matching path interpolates between Gaussian noise and the target radar field, with the model learning the corresponding satellite-conditioned transport velocity.The training objective regresses the network velocity toward the target transport direction.
- Pixel-Space Conditional Flow Matching: Inference starts from Gaussian noise and uses Euler integration over discrete flow-time steps to generate the radar-like field.The sampling schedule uses K steps with Δt = 1/K.
D. Spectral-Local-Wavelet Operator Block
The spectral-local-wavelet operator block combines complementary global, local, and high-frequency processing branches, then adaptively fuses and time-modulates their features for pixel-space flow matching.
- SLW design: The SLW block couples spectral, local, and wavelet operators to model broad organization, convective interactions, and sharp precipitation boundaries.Its residual formulation uses branch-specific operators with input-dependent weights and flow-time modulation.
- Spectral branch: The Fourier spectral branch captures broad precipitation organization and long-range dependencies through low-frequency Fourier convolution.It uses a channel bottleneck and is enabled at lower-resolution encoder levels for more cost-effective global mixing.
- Local branch: The depthwise separable local branch provides short-range spatial mixing for compact convective cells and boundary adjustments without dense convolutions.It combines depthwise 3 × 3 convolution with pointwise projection.
- Wavelet branch: The Haar wavelet branch shrinks unstable high-frequency responses while retaining localized radar texture and sharp echo boundaries.A learned channel-wise threshold is applied to the three high-frequency components before reconstruction.
- Adaptive fusion: Input-dependent softmax gates select the mixture of spectral, local, and wavelet branches for each feature map.The fused feature is further modulated by flow time using feature-wise scale and shift terms.
E. Lightweight Encoder-Decoder
The lightweight encoder-decoder reduces velocity-network cost with narrow channel widths, depthwise processing, and gated additive skip fusion while preserving encoder details.
- Skip fusion: Gated additive skip fusion replaces conventional high-channel skip concatenation, controlling encoder-detail injection without increasing decoder channel width.The skip gate predicts detail injection during upsampling, complementing the compact SLW feature transformations.
- Architecture: The encoder-decoder uses channel multipliers (1, 2, 4) and base width 40, with depthwise downsampling and two SLW blocks at the bottleneck.The bottleneck places spectral mixing where it is cheapest and has the largest effective field of view.
F. Large-Area Tiled Inference Protocol
WaveOp-LiteFM supports large-area retrieval through convolutional inference on overlapping tiles and reports benchmark comparisons spanning efficiency, reconstruction quality, and threshold skill.
- Tiled inference: Overlapping tile predictions are assembled into continuous retrieval fields using Hann-window weighted stitching for China-scale inference.The convolutional spatial operators support flexible grid sizes within memory limits, without fixed-length token sequences or learned absolute positional embeddings.
- Evaluation scope: WaveOp-LiteFM is evaluated on southeast China satellite-to-reflectivity and SEVIR satellite-to-VIL benchmarks, with additional large-area inference over China.The evaluation includes image quality, threshold skill, qualitative cases, model size, FLOPs, and runtime.
- Benchmark comparison: WaveOp-LiteFM maintains strong high-threshold skill while preserving a compact efficiency profile on SE China and SEVIR.Fig. 4 compares efficiency, reconstruction quality, and average versus high-intensity event detection using direction-aware normalized metrics.
- Qualitative retrieval: The model better preserves compact precipitation cores and sharp echo boundaries than baselines in representative SEVIR storm cases.The qualitative panels compare predictions and absolute-error maps for spatial displacement, intensity underestimation, and false precipitation spread.
C. Efficiency
WaveOp-LiteFM reduces the computational cost of pixel-space flow matching while preserving retrieval quality across benchmark and deployment-oriented evaluations. Its compact operator design supports efficient large-area inference and strong ablation performance.
- Efficiency: 52.9% fewer trainable parameters and about 6.7× lower sampling FLOPs than LiteFM-UNet are achieved on both datasets.The comparison uses the same pixel-space flow objective and sampler, isolating the velocity-backbone replacement.
- Large-area inference: Overlapping tiled inference with Hann-window stitching produces broadly consistent large-scale precipitation organization in the China-scale Typhoon Bavi case.The case is used as a qualitative deployment check rather than a pixel-wise benchmark because the retrieval and IMERG reference differ in sensing physics and product definition.
- Ablation performance: The complete model achieves the strongest threshold skill and best SSIM among SEVIR ablations while retaining a compact sampling profile.The wavelet branch is most important for high-frequency details and intense-core skill, while adaptive gates improve robustness across precipitation morphologies.
- Benchmark performance: WaveOp-LiteFM achieves strong radar retrieval skill and perceptual quality with lower sampling cost on SEVIR and southeast China benchmarks.The conclusion reports the cross-dataset outcome without restricting it to a single metric.
APPENDIX
The appendix documents datasets, baselines, training and sampling settings, evaluation conventions, and diagnostic resources for WaveOp-LiteFM. It also describes the study regions and the reported metric directions.
- Datasets and baselines: The experiments use a southeast China FY-4B infrared-to-reflectivity benchmark and the SEVIR visible, infrared, and lightning-to-VIL benchmark.The datasets differ in condition channels, target product, spatial resolution, and evaluation scale.
- Datasets and baselines: LiteFM-UNet shares WaveOp-LiteFM’s objective, 20-step Euler sampler, data splits, and evaluation code but uses a conventional U-Net velocity backbone.This controlled comparison isolates the effect of replacing the velocity network.
- Configuration: WaveOp-LiteFM uses base width 40, channel multipliers (1, 2, 4), a 192-dimensional flow-time embedding, and gated branch and skip fusion.These settings are used for reported runs on both datasets.
- Training and sampling: Inference uses deterministic 20-step Euler sampling after linear interpolation between Gaussian noise and the radar target.The flow-time variable is clipped with ϵ = 10^-4 for numerical stability, with no additional interpolation noise.
- Evaluation protocol: Evaluation reports image-quality metrics, threshold contingency scores, model size, FLOPs estimates, and wall-clock inference time on fixed held-out splits.Southeast China uses thresholds {10, 20, 25, 30, 35}, while SEVIR uses {16, 74, 133, 160, 181, 219}.
D. Diagnostic Figures
The diagnostic figures provide relative multi-metric, baseline, threshold-skill, efficiency, and ablation views that complement exact numerical tables. They are intended to test whether conclusions remain consistent across metrics and architectural components.
- Multi-metric profiles: Figure 10 normalizes quality and cost axes within each dataset to show whether models are balanced across reconstruction, threshold skill, and computational cost.Cost-oriented axes are direction-adjusted, so the radar view is a relative diagnostic rather than a source of exact values.
- Broader baseline comparison: Figure 11 retains raw cell values while normalizing colors within metric columns to compare external baselines, LiteFM-UNet, and WaveOp-LiteFM.The heatmap spans image-level accuracy, threshold skill, perceptual quality, and parameter count.
- Threshold skill: Figure 12 shows raw CSI across evaluated thresholds, revealing how skill changes from light precipitation to intense convective structures.This complements an averaged threshold score and emphasizes sparse, localized echoes that are sensitive to smoothing and displacement.
- Quality-efficiency trade-off: Figure 13 compares RMSE with per-sample computational cost and uses marker size for parameter count to expose quality-efficiency trade-offs.The plot reflects repeated velocity-network evaluations required by pixel-space flow matching.
- Ablation analysis: Figure 14 compares full and ablated architectures under fixed training and sampling protocols to connect changes in quality, threshold skill, and efficiency to components.The diagnostic plots should be read together with exact appendix tables rather than replacing them.
E. Southeast China Qualitative Cases
The qualitative cases examine spatial organization and failure modes that average metrics may miss, while the China-scale example tests tiled deployment beyond the training crop. Together they provide case-level and large-area evidence for retrieval behavior.
- Case-level comparison: Southeast China examples compare satellite conditions, reflectivity-like targets, predictions, and absolute-error maps on the native 500 × 500 regional grid.The cases inspect convective-core displacement, locally intense precipitation underestimation, and false precipitation spread.
- Case-level comparison: High-threshold CSI and POD are sensitive to small spatial shifts in compact convective cells, complementing the qualitative failure-mode inspection.This links case-level evidence to threshold-based evaluation.
- Large-area inference: The Typhoon Bavi case applies overlapping tiled inference across China and compares the stitched retrieval with same-time GPM IMERG precipitation.The example tests visually obvious stitching artifacts and broader spatial deployment rather than pixel-wise accuracy.
G. Detailed Threshold and Efficiency Results
The evaluation isolates WaveOp-LiteFM’s architectural choices and reports substantial efficiency gains for the full model. Ablations assess the spectral branch, wavelet thresholding, and gated fusion across quality, threshold skill, and efficiency.
- Efficiency results: 10116.6 GFLOPs/sample versus 1500.9 for WaveOp-LiteFM on southeast China, while parameters decrease from 5.54M to 2.61M.Both models use 20-step sampling; the comparison is intended to test whether the lightweight backbone preserves sparse intense precipitation skill while reducing computation.
- Efficiency results: 664.8 versus 99.2 GFLOPs/sample on SEVIR, showing lower 20-step sampling cost for WaveOp-LiteFM.
- Ablation study: Removing the spectral branch tests whether low-frequency operator mixing is necessary under the SEVIR setting.
- Ablation study: Removing the wavelet threshold branch tests the contribution of explicit high-frequency shrinkage.
- Ablation study: Disabling gated fusion tests whether input-adaptive branch and skip mixing improves the quality-efficiency trade-off.All ablation variants retain the same optimizer, data split, flow objective, and sampler as the full model.
I. Tiled Large-Area Inference Protocol
WaveOp-LiteFM supports large-area retrieval by applying the fully convolutional model to overlapping tiles and stitching the predictions with weighted blending. The China-scale example uses a broad FY-4B grid and compares the stitched retrieval with a same-time precipitation reference, while the procedure is not a claim of distribution-free generalization.
- Large-area setup: 975 × 1625 grid at 0.04° spacing covers 70.0°E–135.0°E and 15.0°N–54.0°N in the FY-4B China-scale case.
- Tiled inference: Overlapping tiled inference followed by stitching handles regions larger than the training crop or GPU memory budget.The model’s Fourier and Haar operations can be evaluated on arbitrary spatial grids within memory limits.
- Patch sampling: A single global noise realization is cropped consistently across patches, avoiding independently changing stochastic initial conditions from tile to tile.Each patch uses the same 20-step Euler flow-matching solver as the main experiments.
- Prediction stitching: Hann-window blending suppresses tile-edge discontinuities during assembly of overlapping predictions.
- Large-area example: The stitched retrieval is compared with two FY-4B infrared inputs and a same-time GPM IMERG precipitation field for Typhoon Bavi.The case illustrates large-scale precipitation organization over China at 00:00 UTC on 11 July 2026.
- Output processing: The normalized stitched output is clipped to [0, 1] and mapped to a reflectivity-like scale for visualization and threshold interpretation.
- Scope boundary: When the target region, season, sensor geometry, or precipitation regime changes, performance should be revalidated with the main threshold and image metrics.The procedure is explicitly a deployment mechanism rather than a claim of distribution-free generalization.