acceptodds
Withdrawn from ICLR 2027

Learning Intrinsic Water-Quality Dynamics with Rainfall for Data-Driven Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Rainfall is an important environmental driver of water-quality variations through processes such as runoff, pollutant transport, dilution, and resuspension. Traditional mechanistic models can explicitly describe these processes but often require substantial process specification and site-specific calibration, limiting their flexibility under changing hydrological conditions. In this work, we explore a data-driven alternative by proposing RaiNet to jointly model multiscale water-quality dynamics and station-specific rainfall effects across relative lags and temporal scales. RaiNet employs LocTrend to capture irregular water-quality dynamics, constructs station-oriented rainfall events from gridded precipitation, and introduces XGateFusion for conditional lag-aware fusion across scales. We further release three real-world multimodal datasets comprising over 150,000 temporally aligned water quality observations and gridded precipitation raster images. Experiments show that RaiNet outperforms general time-series, water quality, diffusion-based, and spatiotemporal models by over 20%, while component-wise analyses confirm the distinct contribution of each module.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Multimodal environmental learning, basin-scale water-quality monitoring, rainfall-water quality interactions

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_3wD3OpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes RaiNet to represent multiscale water-quality dynamics and station-specific rainfall effects for modeling water-quality variations through processes such as runoff, pollutant transport, dilution, and resuspension.

Critical strengths
  1. The focus on a data-driven approach addresses critical gaps in operational water quality management.
Critical weaknesses
  1. The proposed method is incremental. Without a clear distinction, it looks more like an application study rather than a methodological breakthrough suitable for ICLR.

  2. The comparison with general time-series models is misleading since they are not developed for a specific domain.

  3. This work focuses on water research. ICLR may not be the right venue.

Questions for authors
  1. There is almost no implementation details. Provide model size, memory footprint, running time, etc. for a better assessment.

  2. How are the results of competing methods obtained?

Reviewer_AGU6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose RaiNet, a network for data-driven water-quality forecasting that combines historical water-quality measurements with gridded rainfall data. RaiNet consists of a station-aligned rainfall-event encoder that constructs station-relevant rainfall representations from local and basin-wide precipitation, LocTrend to separate evolving background from dynamic deviations without a fixed period, and XGateFusion to apply rainfall-conditioned corrections to the forecasts. A complementary predictor also incorporates basin-wide rainfall and station topology. The authors additionally introduce three multimodal datasets containing over 150,000 aligned time steps from 25 stations across 3 basins. RaiNet achieves strong results against a broad set of forecasting baselines.

Critical strengths
  1. Meaningful multimodal forecasting task and dataset contribution. The paper addresses a relevant environmental problem by combining historical water-quality measurements with gridded rainfall data to model station-specific rainfall effects on water quality. The authors also introduce three multimodal datasets with aligned water-quality observations and gridded rainfall information.
  2. Strong experimental results and ablation studies. RaiNet achieves strong forecasting performance against a broad range of baselines. The authors provide ablation studies across all three basins, offering evidence of the contributions of the proposed architectural components.
Critical weaknesses
  1. Limitations in the experimental evaluation. Although the paper compares against a broad range of baselines, it omits several relevant state-of-the-art forecasting methods discussed in the related work, such as TimeMixer, TimesNet, and FEDformer. Moreover, while some baselines incorporate rainfall information, it is unclear whether the comparisons adequately isolate the benefits of the proposed architecture from those of using additional environmental inputs. Finally, the evaluation relies mainly on MSE and MAE, which may not fully capture performance during extreme rainfall or water-quality events. Event-specific metrics would better demonstrate the approach's practical benefits.
  2. Architectural complexity and limited evidence of generalization. RaiNet combines several specialized components that progressively construct and refine the forecasts, resulting in a relatively complex architecture. Although the ablation studies provide evidence for the individual components, it remains unclear whether this complexity is necessary compared with simpler multimodal approaches. Furthermore, the paper discusses the limited transferability and site-specific calibration of mechanistic models as motivation, but RaiNet is trained separately for each basin, without evaluating cross-basin transferability or adaptation to unseen stations. Consequently, it is difficult to assess whether the proposed approach offers greater flexibility or generalization beyond the specific settings considered.
Questions for authors
  1. How does RaiNet perform compared with other relevant forecasting methods discussed in the related work, such as TimeMixer, TimesNet, and FEDformer? Can the authors also discuss the evaluated baselines more thoroughly in the related work?
  2. Which environmental inputs are used by each baseline, and how does RaiNet perform against simpler rainfall-aware architectures using the same inputs?
  3. How does RaiNet perform during extreme rainfall and water-quality events, considering metrics beyond aggregate MSE and MAE?
  4. How well does RaiNet transfer across basins or adapt to unseen monitoring stations, given that separate models are trained for each basin and the limitations of site-specific calibration and transferability are part of the motivation?
Reviewer_L3PUOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes RaiNet for rainfall-informed water-quality forecasting. It consists of three main components: a station-aligned rainfall event encoder EI\mathcal{E}_I, which extracts station-relevant events through basin-to-local rainfall contrast and temporal aggregation; LocTrend, which separates locally evolving backgrounds from dynamic deviations using a structural dictionary and multiscale modeling; and XGateFusion, which performs conditional rainfall fusion through bidirectional lag associations and dynamic gating in the prediction space. The model adopts a station-wise primary prediction pathway and incorporates basin-level and topological information through a joint predictor. Across three basins, RaiNet achieves the best MSE in all 15 settings and the best MAE in 14 settings.

Critical strengths

1. Well-motivated, task-specific architecture. Multiscale LocTrend and asymmetric cross-scale mixing address the irregular, non-periodic nature of water-quality dynamics. Station-wise modeling and spatial rainfall priors reduce reliance on long-range cross-station propagation, while rainfall event encoding and XGateFusion account for rainfall persistence, relative lags, and scale-dependent effects. The overall architecture is well aligned with the characteristics of the forecasting task.

2. Competitive experimental performance. RaiNet demonstrates consistent performance across water-quality indicators, basins, and forecasting horizons, with particularly strong MSE results. Comparisons against diverse baselines, component ablations, and multi-seed experiments support the effectiveness of the overall framework.

Critical weaknesses

1. Insufficient validation of key designs in LocTrend and cross-scale mixing. The two-stage RBF-kernel initialization K(0)K^{(0)} is not compared with random or K-means initialization, leaving its practical necessity unclear. Although Figure 6 evaluates the contribution of the background-deviation branches, it does not examine alternative propagation directions (reversed or identical directions) or linear-only mixing. Consequently, the independent contributions of asymmetric propagation and multi-kernel mixing remain insufficiently established.

2. Missing targeted ablations for XGateFusion. The paper does not quantitatively compare (i) aperiodic zero-padded FFT with periodic circular FFT, (ii) bidirectional with unidirectional cross-correlation, or (iii) prediction-space with latent-space interactions. Existing module-level and gating ablations cannot isolate the benefits of these design choices, making it difficult to assess their necessity and whether the additional complexity is justified.

Questions for authors

1. MSE/MAE trade-off. Table 10 shows that rainfall event discretization improves MSE but degrades MAE on 170900, while slightly degrading both metrics on 071200. Are these differences primarily driven by a trade-off between strong rainfall events and routine conditions? Could the authors characterize the gains and losses across event intensities or magnitudes of water-quality changes?

2. Rainfall-specific lag associations. XGateFusion computes cross-correlations between y^T(l)\hat y_T^{(l)} and y^T(l)+y^I(l)\hat y_T^{(l)}+\hat y_I^{(l)}, which share the same water-quality baseline. How can the authors establish that the learned lag associations substantially reflect information introduced by rainfall corrections rather than the temporal structure of the shared baseline? How would the lag associations and forecasting performance change if rainfall corrections were removed or shuffled while keeping the baseline fixed?

If the paper is accepted, we encourage the authors to provide direct access links and more comprehensive metadata for the three multimodal datasets to facilitate reproducibility and future research.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.