Forecasting Coastal ENSO Warming in the Niño 1+2 Region Using ConvLSTM: Toward Improved Early Warning in Peru and Ecuador

Abstract

Accurate forecasting of the El Niño‐Southern Oscillation (ENSO) is essential for improving regional climate resilience and managing water‐related risks. While most deep learning studies have focused on the Niño 3.4 region, the Niño 1+2 region, closely linked to extreme coastal warming associated with ENSO that impacts water infrastructure, flood risk, and agriculture in Peru and Ecuador, remains underexplored. This study develops a spatiotemporal Convolutional Long Short‐Term Memory (ConvLSTM) model to forecast sea surface temperature anomalies (SSTA) at lead times of up to 6 months over the tropical Pacific, with evaluation focused on Niño 1+2 and Niño 3.4. The model is trained using monthly ERSSTv5 sea surface temperature (SST) fields spanning 1854–1996 (with validation over 1997–2013 and testing over 2014–2025) and is assessed using field‐based verification (pattern correlation and spatial error metrics), regional indices, and probabilistic diagnostics from a Monte Carlo (MC) Dropout ConvLSTM ensemble. Across major warm events (e.g., 1997–1998, 2015–2016, and 2023), the model reproduces the spatial evolution of the warm tongue and provides coherent regional forecasts, with uncertainty increasing with lead time and largest in Niño 1+2. A targeted comparison with operational dynamical forecast models indicates that the ConvLSTM framework can provide complementary guidance in the El Niño 1+2 region, during rapidly evolving coastal conditions, together with uncertainty intervals that contextualize forecast confidence. By enhancing early detection of coastal warming, this regionally focused deep learning approach provides actionable forecasts to inform national early warning systems, support seasonal water resource planning, optimize infrastructure operations, and strengthen disaster preparedness in climate‐sensitive regions of coastal South America.