Assessing the environmental costs of multi-scale recurrent neural networks for sustainable extreme rainfall nowcasting – Scientific Reports

Accurately anticipating short‑term cloudbursts can be the difference between routine disruptions and life‑threatening floods. Yet the tools most widely used to predict the weather—physics‑based Numerical Weather Prediction (NWP) systems—often falter at the very extremes that matter most for communities, particularly in the Global South. A new analysis focuses on a deep learning alternative, the Multi‑Scale Recurrent Neural Network (MS‑RNN), and examines whether it can deliver timely and precise nowcasts of intense rainfall while also trimming the environmental cost of computation.

Why nowcasting extreme rainfall remains hard

Nowcasting targets the next minutes to a few hours, where the evolution of storms is driven by rapidly changing, localized processes. NWP models, powerful for longer‑range forecasts, can be too coarse and too slow to capture the explosive growth of convective systems. As a result, the most damaging downpours often arrive with limited warning or mislocated intensity, deepening inequalities where early‑warning infrastructure and emergency services are already stretched.

Deep learning’s promise—and its footprint

Deep learning has shown that data‑driven models can learn radar echo patterns that precede hazardous rainfall, sometimes outperforming traditional methods. But there is a catch: many architectures rely on vast parameter counts and heavy compute budgets, which translate into higher energy use, greenhouse‑gas emissions, and water consumption for cooling. The field increasingly recognizes that model accuracy alone is not enough; the environmental costs of training and inference must be part of the equation, especially where electricity is expensive or unreliable.

Enter MS‑RNN: learning across scales

The MS‑RNN framework is designed to represent atmospheric dynamics at multiple spatial and temporal scales. By nesting recurrent units that each specialize in different resolutions, the architecture aims to capture the fine‑scale motion of storm cells while retaining context from broader circulation patterns. The promise is twofold: improve the fidelity of nowcasts and avoid the parameter explosion typical of large, monolithic networks.

Putting the method to the test on real radar

While the approach has been examined in controlled settings, this study applies MS‑RNN to operational weather radar archives from two contrasting environments: TAASRAD19 and the Rio de Janeiro region. These datasets present distinct convective regimes, offering a rigorous check on the model’s ability to generalize. The evaluation zeroes in on challenging scenarios—short‑lead forecasts of heavy precipitation—where both timeliness and precision are critical.

Measuring performance with sustainability in mind

The analysis assesses predictive skill alongside computational efficiency, foregrounding the often‑overlooked environmental implications of AI. Although comprehensive life‑cycle metrics—such as exact energy draw, carbon emissions, or water usage—are rarely reported for nowcasting studies and were not fully quantified here, compute efficiency serves as a pragmatic proxy. Less processing time and smaller memory footprints typically correspond to lower energy use and, by extension, reduced environmental impact.

What the results suggest

The findings indicate that MS‑RNN can sustain strong predictive performance for extreme rainfall events while substantially improving computational efficiency when applied to real‑world radar data. In practice, that means faster runtimes and lighter hardware demands without a meaningful sacrifice in forecast quality. For meteorological services with constrained budgets and limited access to high‑end GPUs, this balance matters: it opens a path to deployable nowcasting that is both effective and less resource‑intensive.

Implications for climate resilience

Reducing the computational load of nowcasting is not just a technical optimization—it is a climate justice issue. Regions most exposed to flash floods and landslides often face the steepest barriers to advanced forecasting, from power instability to equipment costs. A model that achieves competitive accuracy while using fewer resources broadens access to real‑time warnings and can help close capability gaps that exacerbate disaster risk.

Toward transparent environmental accounting

This work underscores a broader need for the weather and climate AI community to adopt standardized reporting of sustainability metrics. Routine disclosure of training and inference energy use, estimated emissions (adjusted for local grid intensity), and water consumption would enable fair comparisons across methods and encourage designs that optimize for both skill and sustainability. Pairing model innovations like MS‑RNN with such transparency can accelerate progress toward greener, more equitable forecasting systems.

What’s next

Future steps include expanding evaluations across additional climates and radar networks, probing performance during the rarest extremes, and integrating direct measurements of energy and water use. Hybrid strategies—blending physics‑based cues with data‑driven inference—could further bolster robustness while keeping computational costs in check. As extreme precipitation intensifies in a warming world, the imperative is clear: make nowcasting smarter, faster, and lighter on the planet.

Bottom line: applying multi‑scale recurrent architectures to real radar observations shows that high‑impact nowcasts need not come with a high environmental price tag. With thoughtful design and transparent accounting, AI can help deliver earlier, more reliable warnings where they are needed most—without overtaxing the very resources communities depend on.