CNN-Based Subnetworks for Improved Modelling of Extreme Flood and Drought Events in the Negro River Basin

Abstract We propose a novel multi-output convolutional neural network (CNN) framework with station-specific subnets to model and analyze historical low- and high-river stages in the Negro River basin, located in northern South America. The basin is largely covered by tropical forest and experiences strong spatial and seasonal variability in rainfall. The study addresses the challenge of reconstructing accurate waterlevel time series in regions with sparse hydrological observations. Using observed data from five gauging stations—Cucuí, Serrinha, Caracaraí, Santa Maria do Boiaçú, and Moura—the study successfully reconstructs historical waterlevel time series. Quantitative evaluation shows that the subnet-based architecture achieves very low errors (MSE $$\lesssim$$ 0.09) and high distance correlation metrics (DC $$\approx$$ 1.00) during the 2021 flood. We also compare modelling results from subnet-based models with those obtained from individually trained station-specific networks, demonstrating that subnetworks more effectively capture both system-level and station-specific hydrological dynamics. The model captures complex temporal patterns, including sudden decreases, gradual recoveries, and flood rises, demonstrating its ability to represent station-specific hydrological dynamics and system-level responses to extreme floods and droughts. The findings highlight the broader implications of the subnet framework for hydrological prediction under climate variability, particularly for improving early-warning systems and operational monitoring in data-scarce basins. By enhancing the reconstruction of extremes and supporting gap-filling and consistency checking, the method contributes to decision-support strategies for managing future flood and drought risks in the Amazon basin. Graphic Abstract This visual summary provides a concise overview of the study’s core findings and methodologies. Water level data were collected from five gauging stations in the Negro River basin, covering regions with strong spatial and seasonal variability in rainfall. The data were preprocessed using normalization and formatting suitable for 1D convolutional layers. The study employed CNN-based subnetwork architectures integrating features from all stations, alongside independently trained CNNs for comparison, to evaluate the benefits of spatially shared learning. The graphical abstract illustrates the model’s capability to accurately reconstruct complex temporal dynamics, including sudden decreases, progressive recoveries, and repiquetes, while demonstrating superior generalization relative to individual networks across heterogeneous hydrological regimes. These results underscore the potential of the approach as a data-driven tool for flood and drought monitoring, gap-filling, and consistency checking in regions with sparse hydrological data, supporting operational applications such as early warning systems and water resource management. By leveraging subnetwork architectures, this study addresses the challenges of monitoring large and hydrologically diverse regions like the Amazon, highlighting the importance of integrative models for capturing basin-scale dynamics.