Physically consistent reconstruction and propagation analysis of droughts using a physics-informed neural network framework

Study region This study focuses on the Pearl River Basin, southern China, a humid yet seasonally variable monsoon basin where energy fluxes, water storage, and human management jointly influence drought onset, duration, and propagation. Study focus Current monitoring frameworks are predominantly empirical and tailored to specific indices, frequently overlooking the fundamental physical laws that maintain process consistency. Focusing on this challenge, this study presents a physics-informed neural network (PINN) that integrates both water-balance and surface energy-balance constraints within a deep learning framework to reconstruct a physically consistent Composite Drought Index (CDI) by fusing multi-source datasets. The index combines these meteorological, land-surface, and hydrological variables into a single latent space, enabling the derivation of drought intensity and evolution through percentile-based classification. New hydrological insights for the region Relative to unconstrained deep-learning baselines, the physics-guided PINN yields improved spatiotemporal coherence with a ∼9–15% reduction in reconstruction error. The research delineates representative drought archetypes, revealing that energy-limited residuals dominate short-term onset while water-storage deficits drive long-term propagation. Furthermore, the CDI demonstrates a critical 1–2 month lead time over traditional meteorological indices, providing an internally consistent representation for spatiotemporal tracking. This framework integrates physical conservation within a data-driven system, establishing a robust foundation for monitoring drought and diagnosing propagation in intricate hydro-climatic settings.