A lightweight artificial-intelligence model for turning satellite observations into radar-like precipitation maps had substantially lower reported sampling cost than a matched U-Net flow model, while recording better scores on several reconstruction and heavy-precipitation measures in two benchmark tests. The trade-off was speed: WaveOp-LiteFM took slightly longer to run in the reported comparisons.
The work, posted as an arXiv preprint, evaluates WaveOp-LiteFM on a southeast China FY-4B infrared-to-reflectivity-like task using a native 500 by 500 grid and on the SEVIR benchmark, which combines visible, infrared and lightning observations to estimate vertically integrated liquid at 128 by 128 resolution. In plain terms, the system learns to infer a radar-style picture of rain from satellite-based inputs.
Matched tests and model design
In a controlled comparison using the same objective, sampler, data splits and evaluation code, WaveOp-LiteFM had 2.61 million parameters, compared with 5.54 million for LiteFM-UNet. Over 20 sampling steps, the reported calculation cost per sample was substantially lower for WaveOp-LiteFM on both benchmarks.
On the southeast China task, WaveOp-LiteFM's mean absolute error was 2.725 versus 3.01 for the baseline, while peak signal-to-noise ratio was 20.31 versus 19.58, structural similarity was 0.578 versus 0.555, and the heavy-precipitation critical success index was 0.146 versus 0.094. The reported comparison showed the same direction for the listed reconstruction and heavy-threshold measures on SEVIR, although the advantage was not uniform across every metric or model comparison.
Under the hood, WaveOp-LiteFM is trained directly in pixel space, with its objective tied to radar intensity and location. Its backbone combines spectral, local and wavelet processing branches, along with adaptive branch gating and gated additive skip fusion in the decoder.
A component comparison
In a SEVIR comparison of the full architecture with variants lacking specific branches or gates, the full model recorded the highest listed high-threshold scores. Its critical success index at the high threshold, CSI@219, was 0.151 and its Heidke skill score, HSS@219, was 0.260; the corresponding scores were 0.132 and 0.238 for the variant without the spectral branch, 0.116 and 0.207 for the variant without the wavelet branch, and 0.114 and 0.203 for the variant without gates.
Across a wider set of comparisons, the authors describe a favorable balance between quality and efficiency against CNN, transformer, GAN, diffusion, flow and weather-specific models. They also report the fewest parameters among the generative baselines.
Compute savings and runtime told different stories
The reported wall-clock measurements complicate the efficiency picture. WaveOp-LiteFM took 429.22 milliseconds per inference on the southeast China setting, versus 416.30 milliseconds for LiteFM-UNet, and 36.03 milliseconds on SEVIR, versus 32.25 milliseconds for the baseline. Hardware and runtime uncertainty estimates are not reported.
The large-area case was qualitative
The researchers also ran the model over a China-scale Typhoon Bavi case by processing overlapping tiles and stitching them with Hann-window weighting. The resulting retrieval broadly matched the large-scale precipitation organization in a GPM IMERG reference, but this was a qualitative check rather than a pixel-by-pixel benchmark because the products use different sensing physics and definitions.
The example does not establish distribution-free transfer across regions, sensors, seasons or precipitation regimes. Tiled inference is described as flexible to grid size within memory limits, but that capability does not by itself show that the reported rankings will hold outside the tested settings.
What the comparisons leave open
The evidence comes from two fixed benchmark evaluations, controlled architectural ablations, computational comparisons and one qualitative large-area case. The supplied report does not provide inferential uncertainty estimates, confidence intervals or statistical significance tests, and the benchmark example counts are not reported.
Broader testing across regions, sensors, seasons and precipitation regimes would be needed to assess transfer. It also remains unclear whether implementation or hardware changes could turn the lower sampling cost into lower wall-clock latency.
Paper data and sources
Original title: WaveOp-LiteFM: Lightweight Neural-Operator Flow Matching for Satellite-to-Radar Precipitation Retrieval
Authors: Chunlei Shi, Yecheng Zhang, Yufeng Zhu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text