🤖 machine learning

From Drops to Grid: Noise-Aware Spatio-Temporal Neural Process for Rainfall Estimation

The paper introduces DropsToGrid, a noise-aware spatio-temporal neural process that fuses sparse, noisy weather station data with radar context to generate accurate, high-resolution rainfall maps with well-calibrated uncertainty, outperforming existing operational and deep learning baselines.

Original authors: Rafael Pablos Sarabia, Joachim Nyborg, Morten Birk, Ira Assent

Published 2026-05-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Rafael Pablos Sarabia, Joachim Nyborg, Morten Birk, Ira Assent

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine trying to paint a perfect, high-resolution picture of rainfall across a whole country, but you only have two very different, imperfect tools to help you.

The Problem: A Messy Puzzle
Think of rainfall like a giant, invisible blanket of water falling from the sky. To understand it, we usually rely on two sources:

The "Crowd" (Private Weather Stations): Imagine thousands of people in their backyards holding rain gauges. Some are very accurate, but many are cheap, broken, or placed in weird spots (like under a tree). They give you specific points of data, but they are scattered like confetti, and some of the numbers are wrong.
The "Eye in the Sky" (Radar): This is like a giant, sweeping flashlight that sees the whole sky at once. It's great for seeing the big picture, but it's not perfect. It measures moisture in the air, not the actual rain hitting the ground, so it often guesses wrong about how much rain is actually falling.

Traditional methods try to connect the dots between these scattered backyard gauges or just trust the radar. But because the rain is messy, local, and the data is noisy, the resulting maps are often blurry, biased, or miss the small, intense storms entirely.

The Solution: DropsToGrid
The authors created a new AI tool called DropsToGrid. Think of it as a super-smart, artistic translator that can take those messy, scattered backyard measurements and the blurry radar images and blend them into a single, crystal-clear, high-definition map of exactly where the rain is falling.

Here is how it works, using some simple analogies:

The "Neural Process" (The Intuitive Artist): Instead of just memorizing patterns, this AI acts like an artist who understands uncertainty. If the backyard gauges are far apart, the artist knows, "I'm not 100% sure what's happening in the middle, so I'll paint a slightly fuzzy area there." If the radar says it's raining but the ground gauges say it's dry, the artist knows to trust the ground gauges more but still uses the radar to understand the shape of the storm. It doesn't just guess; it tells you how confident it is in its guess.
The "Translation Equivariance" (The Moving Puzzle): Rain doesn't stay still; it moves. If a storm moves 10 miles to the east, the AI's prediction should move 10 miles to the east, too. This tool is built like a sliding puzzle that understands movement. It doesn't matter where the storm is; the AI knows how to handle it the same way, so it works well even in new cities it has never seen before.
The "Zero-Inflated Gamma" (The Rain-or-Nothing Logic): Rain is weird. Most of the time, it's not raining at all (zero). When it does rain, it can be a drizzle or a flood. Standard math struggles with this "all or nothing" nature. This AI uses a special mathematical trick (Zero-Inflated Gamma) that treats "no rain" as a distinct event from "light rain" or "heavy rain," allowing it to model the reality of weather much better than old methods.

What They Found
The team tested this tool on real data from Denmark and then across all of Europe.

Better than the Pros: It beat the standard, professional weather models used by governments and the other top AI models currently available.
Works with Fewer Gauges: Even if you only have 5% of the backyard gauges working, DropsToGrid can still create a much better map than the old radar-only methods.
Knows What It Doesn't Know: The tool produces a "confidence map" alongside the rain map. It highlights areas where it's unsure (usually where there are no gauges), which is incredibly useful for knowing where to put new sensors.

In a Nutshell
DropsToGrid is a new way to turn scattered, noisy, and imperfect rain reports from the ground and the sky into a reliable, high-definition picture of the weather. It's like taking a blurry, low-res photo and using a smart filter to sharpen it, while also telling you exactly which parts of the photo are still a bit fuzzy.

Technical Summary: From Drops to Grid: Noise-Aware Spatio-Temporal Neural Process for Rainfall Estimation

1. Problem Statement

Accurate, high-resolution rainfall estimation is critical for weather forecasting, water management, and hazard mitigation. However, existing data sources present significant limitations:

Ground Gauges (SYNOP): While physically accurate, national networks are too sparse to capture microscale (<100m) or local-scale (100m–3km) variability.
Remote Sensing (Radar/Satellite): Radar provides dense coverage but suffers from biases due to beam blocking, attenuation, and the need for error-prone empirical conversions from reflectivity to rainfall. Satellite estimates often lack resolution or reliability in specific regions.
Private Weather Stations (PWS): Crowd-sourced PWS networks offer high density but are characterized by irregular spatial distribution, temporal inconsistency (stations appearing/disappearing), and significant noise due to varying hardware quality and installation.

The core challenge is to densify rainfall fields by fusing sparse, noisy, and irregularly distributed PWS measurements with spatially continuous but biased radar data. This requires a model capable of spatio-temporal reasoning, uncertainty quantification, and multi-modal fusion that can handle non-pixel-aligned, evolving inputs without overfitting to noise.

2. Methodology: DropsToGrid

The authors propose DropsToGrid, a Neural Process (NP)-based method designed to generate stochastic, continuous rainfall fields with calibrated uncertainty. The architecture integrates temporal station sequences with radar-derived spatial context.

2.1. Core Architecture

DropsToGrid extends the Convolutional Conditional Neural Process (ConvCNP) framework to handle heterogeneous, non-pixel-aligned sources.

Translation Equivariance: The model is designed to be translation-equivariant, ensuring that spatial shifts in input produce equivalent shifts in output. This allows for generalization across regions without retraining for specific station layouts.
Sparse-to-Dense Encoding (SetConv):
- Irregular PWS measurements and radar data are processed through independent SetConv encoders.
- SetConv projects sparse observations into a continuous latent field using a density channel to normalize overlapping contributions, effectively translating sparse inputs into smooth, dense representations.
- Modality-specific MLPs embed these dense representations into feature vectors.
Spatio-Temporal Bottleneck Fusion:
- Temporal Encoder: A Temporal Pixelwise Transformer with a learned global query aggregates information across time steps for station data, capturing short- and long-term dependencies.
- Gated Cross-Attention: A Translation-Equivariant (TE) Transformer performs cross-modal reasoning. Station features act as queries, and radar features as keys/values. A learned spatial gate modulates radar contributions, allowing the model to selectively attend to radar regions relevant to ground observations while suppressing misaligned or uninformative areas.
- TE-Fusion Transformer: The final fusion layer enforces translation equivariance by computing attention weights based on relative displacements rather than absolute positions, ensuring scale invariance and spatial coherence.
Probabilistic Decoder: A U-Net decoder upsamples the fused latent features to full resolution, outputting parameters for a probabilistic distribution rather than a single deterministic value.

2.2. Noise-Aware Objective (ZIG)

Recognizing that rainfall data is zero-inflated (many dry periods) and heavily right-skewed, the model employs a Zero-Inflated Gamma (ZIG) likelihood.

The distribution models two components: a Bernoulli distribution for the probability of zero rainfall ( $\pi_0$ ) and a Gamma distribution for positive rainfall intensities.
The loss function maximizes the log-likelihood of the ZIG distribution, explicitly separating the probability of no rain from the intensity of wet conditions.
To prevent overfitting to noisy PWS inputs, the loss is computed only on target grid cells that do not contain input station data during training, encouraging generalization to unseen locations.

3. Key Contributions

First NP-based Rainfall Densification: Introduces the first Neural Process approach for densifying rainfall from PWS data, combining temporal sequences and radar context in a single probabilistic framework.
Multi-Modal Attention Mechanism: Proposes a translation-equivariant fusion architecture that effectively aligns sparse, noisy ground observations with dense radar data using gated cross-attention and relative-position-based transformers.
Comprehensive Empirical Evaluation: Conducts a thorough study including ablation analyses, station density assessments, uncertainty quantification, and cross-regional evaluations across Europe, comparing against both operational baselines and deep learning models.

4. Experimental Results

The model was evaluated on real-world datasets covering Denmark (2024) and extended to a Europe-wide evaluation (2025).

Performance vs. Operational Baselines: DropsToGrid outperformed all operational estimators (OPERA, RainViewer, IMERG, ERA5, and DMI Climate) across all metrics.
- CSI (Critical Success Index): Achieved 0.532 on PWS holdout and 0.551 on SYNOP stations, significantly higher than the next best (Climate at 0.437 and OPERA at 0.298).
- FSS (Fraction Skill Score): Demonstrated superior spatial structure accuracy (0.819 vs. 0.740 for Climate).
- FBI (Frequency Bias Index): Achieved a bias closest to 1 (0.877), indicating minimal systematic over- or under-prediction compared to other models.
Performance vs. Deep Learning Baselines: Outperformed ConvCNP and SwinTNP variants (both standard and translation-equivariant) across CSI, FSS, and CRPS (Continuous Ranked Probability Score).
Robustness to Sparsity: Even when using only 5% of available PWS stations, DropsToGrid outperformed operational estimators by over 24% in CSI, demonstrating effectiveness in low-density regions.
Cross-Regional Generalization: A model trained solely on Danish data successfully generalized to the entire European continent in 2025 (unseen during training), maintaining robust performance despite lower data quality and sparser coverage in the EU compared to Denmark.
Uncertainty Calibration: The model produces well-calibrated uncertainty maps, with higher uncertainty in regions farther from stations or where rainfall is present, reflecting awareness of observation noise.

5. Significance and Claims

The paper claims that DropsToGrid represents a significant advancement in rainfall estimation by addressing the specific challenges of noise, irregularity, and sparsity inherent in crowd-sourced weather data.

Reliability: It generates accurate high-resolution rainfall maps with calibrated uncertainty, even when few stations are available.
Generalization: The translation-equivariant design allows the model to generalize across different regions and temporal scales without retraining for specific station layouts.
Practical Utility: The uncertainty-aware framework provides actionable insights for optimizing sensor placement and assessing data quality.
Scalability: The method is computationally efficient (inference takes 0.20s for EU-wide patches vs. 0.47s for baselines) and scalable to large geographic areas.

The authors position DropsToGrid not just as a densification tool, but as a step toward a general probabilistic data assimilation framework that can integrate various meteorological variables beyond just rainfall.