Station2Radar: query conditioned gaussian splatting for precipitation field

Doyi Kim; Minseok Seo; Changick Kim

cs.CVFeb 28, 2026

Station2Radar: query conditioned gaussian splatting for precipitation field

Authors: Doyi Kim, Minseok Seo, Changick Kim

Organizations: Korea Advanced Institute of Science and Technology (KAIST)

Abstract

Precipitation forecasting relies on heterogeneous data. Weather radar is accurate, but coverage is geographically limited and costly to maintain. Weather stations provide accurate but sparse point measurements, while satellites offer dense, high-resolution coverage without direct rainfall retrieval. To overcome these limitations, we propose Query-Conditioned Gaussian Splatting (QCGS), the first framework to fuse automatic weather station (AWS) observations with satellite imagery for generating precipitation fields. Unlike conventional 2D Gaussian splatting, which renders the entire image plane, QCGS selectively renders only queried precipitation regions, avoiding unnecessary computation in non-precipitating areas while preserving sharp precipitation structures. The framework combines a radar point proposal network that identifies rainfall-support locations with an implicit neural representation (INR) network that predicts Gaussian parameters for each point. QCGS enables efficient, resolution-flexible precipitation field generation in real time. Through extensive evaluation with benchmark precipitation products, QCGS demonstrates over 50% improvement in RMSE compared to conventional gridded precipitation products, and consistently maintains high performance across multiple spatiotemporal scales.

Figures & tables

Figure 1: Construction of precipitation fields from (a) sparse AWS observations. (b) Empirical kernel-based interpolation oversmooths and blurs rainfall boundaries. (c) QCGS leverages satellite context and gauge anchors to selectively place Gaussians, producing resolution-flexible and structurally consistent fields. (d) Ground-truth radar at 2 km resolution for reference.

Figure 2: Overview of the proposed QCGS pipeline. AWS observations and satellite BT imagery are fused to produce a coarse surrogate field and candidate rainfall-support points. A rainfall-aware sampling strategy and an INR-based Gaussian estimator then predict splatting parameters, yielding resolution-flexible precipitation fields through selective Gaussian rendering.

Figure 3: Visualization of different point sampling strategies for precipitation fields. (a) Ground truth radar field, (b) uniform sampling, (c) edge-based sampling, (d) heavy-rain sampling, and (e) our mixed strategy. Uniform sampling provides overall coverage but lacks details in heavy rainfall regions. Edge-based sampling emphasizes boundaries but overlooks fine-scale structure. Heavy-rain sampling concentrates points on intense precipitation, leaving light-rain areas underrepresented. In contrast, our mixed strategy balances intensity-aware sampling with spatial coverage across the field.

Temporal scale (cases)	Category	Method	Res.	RMSE $\downarrow$	LPIPS $\downarrow$	CSI $\uparrow$	FSS $\uparrow$	CC $\uparrow$	Bias $\approx$ 1
Snapshot (1154)	Data-driven	Pix2PixHD	0.5 km	2.45	0.62	0.59	0.71	0.55	0.82
	Data-driven	NPM	0.5 km	1.95	0.58	0.59	0.78	0.68	0.88
	Data-driven	BBDM	0.5 km	1.68	0.54	0.64	0.84	0.75	0.93
	Satellite Product	GK2A	2.0 km	2.89	0.40	0.20	0.37	0.12	-
	Classical Interp.	Barnes	2.0 km	2.56	0.39	0.47	0.68	0.42	0.98
	Classical Interp.	Kriging	2.0 km	2.43	0.40	0.50	0.69	0.45	1.03

Table 1: Quantitative results across multiple spatiotemporal scales. Each block shows the number of evaluated cases in parentheses. QCGS is trained at 2 km and downsampled to

0.1^{\circ}

for comparison with global products. Best scores per block are in bold .

Figure 4: Panels (a)–(e) show the comparison of daily accumulated rainfall (mm, day) between radar and four rainfall products: IMERG, GSMaP, MSWEP, and QCGS. The “Total diff.” values (mm, day) denote the spatially integrated difference from radar for each product. Panel (f) presents the PSD across spatial scales (since $1^{\circ}\approx 111$ km, a wavenumber of 1(1/degree) corresponds approximately to a spatial scale of $\sim$ 110 km.)

Architecture	CSI $\uparrow$
AWS (only)	0.53
U-Net (ConvNeXt)	0.62
+ AWS fusion	0.73
+ AWS fusion + GS (ours)	0.76

Table 2: Comprehensive ablation study on CSI. (a) Effect of architecture choices: AWS fusion and Gaussian Splatting (GS) progressively improve performance. (b) Effect of sampling strategy: combining gradient, regular, and heavy-rain sampling yields the best CSI. (c) Effect of the number of query points: performance saturates around

K{=}6000

. Best results are in bold .

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Examples of consecutive one-hour QCGS-generated frames. The last three frames exhibit rainfall patterns that were absent in the first three, indicating temporal inconsistencies. Such frame-to-frame mismatch can hinder the performance of video prediction models that rely on coherent temporal dynamics.

Model	CSI@1mm (R $\rightarrow$ R)	CSI@1mm (QCGS $\rightarrow$ R)
MetNet-v2 Sønderby et al. (2020)	0.390	0.374
SimVP Gao et al. (2022a)	0.591	0.252
PreDiff Gao et al. (2023)	0.664	0.381

Appendix

Table 3: Forecasting performance at +60 minutes using QCGS-generated radar fields as inputs. Baselines were trained only on radar-to-radar data and used without retraining.

Figure 6: Bias, RMSE, and correlation coefficient (CC) of AWS rain rate compared with QCGS and radar. QCGS consistently achieves lower bias and RMSE and higher CC relative to radar, demonstrating closer agreement with gauge observations.

Figure 7: Case studies comparing QCGS and radar against AWS stations. Gridded fields are spatially matched to AWS locations, showing that QCGS preserves local rainfall intensities more faithfully than radar.

Threshold	QCGS (0.5 km)	QCGS (2 km)	QCGS (10 km)	IMERG	GSMaP
1	0.657	0.703	0.506	0.366	0.308
5	0.415	0.483	0.306	0.140	0.129
10	0.311	0.401	0.232	0.046	0.065

Appendix

Table 4: CSI scores at different rainfall thresholds (mm/hour) using hourly data. QCGS is evaluated at multiple spatial resolutions (0.5, 2, and 10 km).

	POD			FAR			CSI
Threshold	10	50	100	10	50	100	10	50	100
QCGS	0.703	0.579	0.646	0.125	0.329	0.423	0.657	0.455	0.434
IMERG	0.679	0.369	0.117	0.267	0.616	0.614	0.541	0.173	0.039
GSMaP	0.591	0.358	0.277	0.262	0.710	0.754	0.493	0.165	0.119
MSWEP	0.714	0.315	0.096	0.286	0.554	0.584	0.553	0.191	0.067

Appendix

Table 5: Categorical POD, FAR, and CSI scores at different rainfall thresholds (mm, day) for daily accumulation data.

Figure 9: Visual comparison of Radar, QCGS, AWS-only, and Satellite-only across different cases.

Figure 10: Qualitative comparison of observations, global rainfall products, and QCGS. QCGS preserves fine-scale precipitation structures and reduces large-scale biases relative to conventional products.

Figure 11: Qualitative comparison between classical interpolation methods and QCGS (AWS-only).

Figure 12: Qualitative comparison of precipitation fields from radar, QCGS, IMERG, and GSMaP. Radar provides the reference, while QCGS preserves fine-scale structures more faithfully than global products. IMERG and GSMaP show smoother fields with biases in convective regions.

Figure 13: Qualitative comparison of precipitation fields from radar, QCGS, IMERG, and GSMaP. Radar provides the reference, while QCGS preserves fine-scale structures more faithfully than global products. IMERG and GSMaP show smoother fields with biases in convective regions.

Gaussian Splatting

Gaussian Splatting represents scenes with learnable Gaussian primitives for fast rendering, 3D reconstruction, and dynamic-scene modeling.

See all 82 papers

Neural Stochastic Processes for Satellite Precipitation Refinement
Apr 12, 2026Shunya Nagashima, Takumi Bannai, Shuitsu Koyama +2Neural Processes Remote Sensing
Neurogica Inc. · LTS, Inc. · ME-Lab Japan, Inc.
FusionCast: Enhancing Precipitation Nowcasting with Asymmetric Cross-Modal Fusion and Future Radar Priors
Mar 2, 2026Henan Wang, Shengwu Xiong, Yifang Zhang +4Precipitation Nowcasting Precipitation Forecasting
School of Computer Science and Artificial Intelligence, Wuhan University of Technology · School of Earth and Space Science and Technology, Wuhan University
A plug-and-play generative framework for multi-satellite precipitation estimation
May 14, 2026Yunfan Yang, Haofei Sun, Xiuyu Sun +7Remote Sensing Generative Modeling
State Key Laboratory of Atmospheric Boundary Layer Physics and Atmospheric Chemistry (LAPC), Institute of Atmospheric Physics, Chinese Academy of Sciences, Beijing, China. · College of Earth and Planetary Sciences, University of Chinese Academy of Sciences, Beijing, China. · Shanghai Academy of Artificial Intelligence for Science (SAIS), Shanghai, China. +5

Apr 12, 2026J/K move · Enter open · S save

Neural Stochastic Processes for Satellite Precipitation Refinement

Shunya Nagashima, Takumi Bannai, Shuitsu Koyama, Tomoya Mitsui+1

Neurogica Inc. · LTS, Inc. · ME-Lab Japan, Inc.

Neural Processes Remote Sensing Weather Forecasting Remote Sensing Benchmarks

Open paper AskPDF arXiv

Accurate precipitation estimation is critical for flood forecasting, water resource management, and disaster preparedness. Satellite products provide global hourly coverage but contain systematic biases; ground-based gauges are accurate at point locations but too sparse for direct gridded correction. Existing methods fuse these sources by interpolating gauge observations onto the satellite grid, but treat each time step independently and therefore discard temporal structure in precipitation fields. We propose Neural Stochastic Process (NSP), a model that pairs a Neural Process encoder conditioning on arbitrary sets of gauge observations with a latent Neural SDE on a 2D spatial representation. NSP is trained under a single variational objective with simulation-free cost. We also introduce QPEBench, a benchmark of 43{,}756 hourly samples over the Contiguous United States (2021--2025) with four aligned data sources and six evaluation metrics. On QPEBench, NSP outperforms 13 baselines across all six metrics and surpasses JAXA's operational gauge-calibrated product. An additional experiment on Kyushu, Japan confirms generalization to a different region with independent data sources.

Abstract

Gaussian Splatting

Explore similar work

Neural Stochastic Processes for Satellite Precipitation Refinement

FusionCast: Enhancing Precipitation Nowcasting with Asymmetric Cross-Modal Fusion and Future Radar Priors

A plug-and-play generative framework for multi-satellite precipitation estimation