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

Lilly Horvath-Makkos, Fayyaz Minhas, Department of Computer Science, The University of Warwick, Coventry, United Kingdom · 2026

Read Interpretable rainfall modelling reveals rapid reorganisation of Amazonian rainfall under vegetation loss by Lilly Horvath-Makkos, Fayyaz Minhas, Department of Computer Science, The University of Warwick, Coventry, United Kingdom (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-10948.txt.

Introduction

Lilly Horvath-Makkos and Fayyaz Minhas model how vegetation loss reorganizes Amazon rainfall, hour by hour. Deforestation shifts rain through tangled, seasonal, nonlinear land-atmosphere feedbacks, yet most models blur convection at coarse scales and resolve correlations rather than causal pathways.

A neural network forecasting hourly rainfall across the Amazon Basin, paired with pathway diagnostics and sensitivity tests, checks whether the model internalizes physically consistent links from canopy to atmosphere to rain. The model predicts occurrence and intensity well, and sustained canopy loss shows threshold-like reorganization in rainfall patterns. Climate and water planners get an interpretable warning system for hydrological disruption.

Catalog record

CREATORLilly Horvath-Makkos, Fayyaz Minhas, Department of Computer Science, The University of Warwick, Coventry, United Kingdom
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE93778
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.10948

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Interpretable rainfall modelling reveals rapid reorganisation of Amazonian rainfall under vegetation loss Lilly Horvath-Makkos, Fayyaz Minhas, Department of Computer Science, The University of Warwick, Coventry, United Kingdom 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. Graphical abstract: Interpretable framework for analysing vegetation-rainfall coupling over the Amazon Basin using ERA5 reanalysis data (2021–2024) at hourly resolution. A MultiTask ConvLSTM predicts rainfall occurrence and intensity from atmospheric and vegetation inputs (Spearman \rho\approx 0.84 , F1 \approx 0.93 , ROC-AUC \approx 0.98 ), enabling sensitivity-based attribution of land-atmosphere interactions consistent with established ecohydrological mechanisms. Counterfactual deforestation experiments reveal nonlinear and asymmetric rainfall responses, including reductions of up to 7% in heavy rainfall (20–50 mm h -1 ), increases of up to 4% in light rainfall (0–1 mm h -1 ), and an abrupt decline in precipitating area fraction after approximately 2–3 months of sustained vegetation loss along the most sensitive directions. Spatial sensitivity mapping identifies the north-western Amazon and Andean foothills as hotspots of short-term hydrological vulnerability. 1 Introduction Tropical forests play a central role in regulating regional and global rainfall, with Amazonian deforestation disrupting hydrological cycles through reduced evapotranspiration, altered convection, and displacement of the intertropical convergence zone 4 ; 3 ; 10 ; 27 ; 31 . These disruptions affect biodiversity, water security, and agriculture 44 ; 32 , and are increasingly linked to tipping points in both regional and global climate 18 ; 50 . Currently, over 20% of the Amazon rainforest is estimated to have been lost 14 . As deforestation shows little sign of slowing to manageable levels, understanding short-term rainfall responses is increasingly urgent for mitigating local and regional climate impacts. Rainfall regulation in tropical forests arises from tightly coupled exchanges of moisture, energy, and momentum between the land surface and atmosphere. Vegetation-related variables capture key pathways of land-atmosphere coupling: Leaf Area Index (LAI) controls evapotranspiration capacity and boundary-layer humidity 22 ; canopy and surface roughness regulate turbulent heat and momentum fluxes that influence boundary-layer depth and convective triggering 37 ; surface albedo alters radiative balance 67 ; and soil water availability constrains latent heat flux, reinforcing drought through suppressed evapotranspiration 55 . Operating across hours to months, these processes complicate attribution, as observed rainfall variability often reflects a combination of causal land-atmosphere feedbacks and coincident atmospheric variability. In the Amazon Basin, deforestation reduces latent heat flux, increases sensible heating, decreases roughness, and increases albedo, collectively suppressing convection and altering moisture convergence 33 . Accurately capturing how these fast, localised processes generate rainfall extremes remains challenging, as many approaches smooth convective structure, misplace precipitation, underpredict high-intensity events, or lack mechanisms for attribution beyond correlation 60 ; 39 ; 34 ; 29 ; 53 ; 28 ; 17 ; 11 ; 61 . Rainfall responses to deforestation are also highly spatially heterogeneous, further limiting inference based on correlation alone. Southern Amazonia consistently exhibits drying in climate model experiments 4 ; 2 , reflecting its distance from Atlantic moisture sources and reliance on local recycling 50 , whereas northwestern Amazonia often shows weaker sensitivity due to persistent Atlantic and Caribbean inflow 45 ; 19 . The Andes further modulate these responses by intensifying convection along the western margin while diverting low-level jets that redistribute moisture southward 12 . Large river systems such as the Amazon and Negro, together with coastal circulations, introduce additional sources of local moisture recycling, producing regions where rainfall may increase locally despite broader regional drying 33 . As a result, spatial correlations between forest loss and rainfall can mask underlying feedbacks and yield opposing trends depending on scale. Sustained deforestation within concentrated hotspots may further induce hysteresis, whereby reduced evapotranspiration and enhanced sensible heating suppress convection beyond a threshold, limiting rainfall recovery even under regrowth 59 . Identifying when and where such feedbacks emerge therefore, requires approaches that move beyond correlation toward causal reasoning at short timescales relevant for land-atmosphere coupling. This will assist in understanding the resilience of rainfall to disturbance. Previous studies using global and regional climate models, including WRF-Noah 4 and OLAM 2 , have documented substantial rainfall reductions in the southern Amazon under deforestation and revealed associated changes in moisture convergence and large-scale circulation. Observational analyses further suggest threshold behaviour, whereby prolonged forest loss in long-established hotspot regions ( > 10 years) leads to significant drying 44 , while satellite-based estimates associate each 1% forest loss with an average decline of 0.25 mm month -1 in rainfall 58 . Together, these studies establish the sensitivity of Amazonian rainfall to vegetation loss, but leave open questions regarding short-term responses, spatially localised effects, and causal attribution. Moreover, numerical weather prediction and climate modelling frameworks typically rely on ensemble simulations and substantial computational resources, operate at coarse spatial or temporal resolution, and often struggle to accurately represent convective precipitation, limiting their ability to resolve fast, localised land-atmosphere feedbacks. Recent advances in neural weather forecasting, including models such as ISOX 36 and 4DVarFormer 61 , have demonstrated strong performance in large-scale precipitation estimation and extended-range forecasting. More broadly, machine learning (ML) approaches have gained prominence for precipitation nowcasting by learning spatiotemporal rainfall patterns directly from high-resolution reanalysis or radar data, complementing process-based climate and land-surface models. In particular, deep learning architectures such as Convolutional Long Short-Term Memory Networks (ConvLSTMs) and attention-based models have shown strong skill in predicting convective rainfall events 61 ; 57 ; 56 ; 20 ; 51 , making them well suited for analysing short-term, localised precipitation dynamics and supporting efficient

Read the complete text