# Precipitation nowcasting with radar data for evaluating multiple horizons using U-Net-based algorithm in Eastern Amazon

## TL;DR

This paper uses a U-Net machine learning model to improve short-term rainfall forecasts in the Eastern Amazon, finding that using more than 60 minutes of past data reduces accuracy.

## Contribution

The study introduces a U-Net-based approach for multi-horizon precipitation nowcasting and identifies optimal input data lengths for accurate forecasts.

## Key findings

- Using more than 60 minutes of input data degrades short-term forecast accuracy.
- 120 minutes of input data led to a 17.60% RMSE increase and 7.18% CSI decrease compared to 60 minutes.
- Optimal input data length improves nowcasting accuracy for severe weather alerts.

## Abstract

Severe meteorological events are increasingly frequent globally, with intense rainfall significantly impacting well-being, safety, and the economy, including agriculture and mining. Timely emergency alerts are crucial for mitigating losses and preventing fatalities from extreme weather. Precipitation forecasting tools, especially meteorological radars and satellites, are vital due to their high temporal resolution. This study utilizes a U-Net machine learning architecture for spatial-temporal precipitation nowcasting. We evaluate a multi-horizon nowcasting approach using meteorological radar data from the Eastern Amazon, investigating the input data (past horizons) needed for optimal forecast horizons. Our results show that increasing input data beyond 60 minutes degrades performance for short forecast horizon. For short-term forecasts, using 120 minutes of input data instead of 60 minutes resulted in a significant performance loss of 17.60% in RMSE and 7.18% in CSI. These findings identify the optimal input data for accurate nowcasting, enabling safer decision-making during severe weather.