Interpretable rainfall modelling reveals rapid reorganisation of Amazonian rainfall under vegetation loss

Abstract. Understanding how vegetation loss alters rainfall remains a major challenge in climate and hydrological science, as deforestation modifies precipitation through heterogeneous, seasonal, and nonlinear land-atmosphere feedbacks. Existing modelling systems struggle to capture these dynamics: convection is parameterised at coarse spatial scales, potential tipping behaviour is poorly constrained, and rainfall–deforestation analyses are often limited to multi-decadal timescales. As a result, many approaches resolve correlations rather than causal effects, limiting our ability to anticipate hydrological disruption and inform water-security planning. Using a neural-network predictive model for hourly rainfall forecasting across the Amazon Basin, coupled with mechanistic pathway diagnostics and sensitivity analyses, we examine how vegetation perturbations reorganise rainfall dynamics across space, intensity regimes, and timescales. We test whether the model internalises physically consistent pathways linking vegetation, atmospheric state, and precipitation, and whether sustained canopy loss is associated with threshold behaviour in rainfall organisation. The model accurately predicts rainfall occurrence and intensity across the Amazon (Spearman = 0.84, F1 = 0.93, ROC-AUC = 0.98) and learns temporally ordered, physically consistent dependencies aligned with ecohydrological theory. From sensitivity analyses, we observe rapid and asymmetric rainfall responses to vegetation loss: heavy rainfall (20–50 mm h^-1) declines by up to 7 % under sustained deforestation over eight months, while light rainfall (0.1–1 mm h^-1) increases by nearly 4 %. Across scenarios, we observe rainfall entropy increases by 1.3 %, and dry-season intensity rises by 0.3–0.5 % per 0.5 % forest-cover loss, with the strongest disruptions occurring in the north-western Amazon and the Andean foothills. Through threshold analysis, we observe that after approximately 2–3 months of sustained vegetation changes in the most sensitive regions, the precipitating area fraction declines sharply. These findings demonstrate that data-driven methods can uncover process-relevant signatures of land–atmosphere coupling, offer new insights into hydrological vulnerability, and emphasise the urgency of Amazon conservation.