🔬 physics

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

By employing an interpretable neural network model, this study demonstrates that sustained Amazonian deforestation triggers rapid, asymmetric reorganization of rainfall patterns—characterized by a sharp decline in heavy rainfall, a rise in light rainfall, and threshold-driven hydrological disruption—thereby revealing previously underappreciated vulnerabilities in the region's land-atmosphere coupling.

Original authors: Lilly Horvath-Makkos, Fayyaz Minhas

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lilly Horvath-Makkos, Fayyaz Minhas

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine the Amazon rainforest not just as a collection of trees, but as a giant, living air conditioner and water pump that keeps the region's weather running smoothly. For a long time, scientists have worried that if we cut down too many trees, this system might break, leading to droughts or strange weather. But figuring out exactly how and when that happens has been like trying to predict the weather while wearing thick foggy glasses.

This paper introduces a new, super-smart computer program (an AI model) that acts like a "weather detective" to solve this mystery. Here is what the researchers found, explained simply:

1. The Detective's Tool: A "Time-Traveling" Weather Model

The researchers built a neural network (a type of AI) that learns from hourly weather data from 2021 to 2024. Think of it as a detective who has watched the Amazon's weather every single hour for three years.

What it does: It predicts whether it will rain and how hard, hour by hour.
How good is it? It's incredibly accurate. If you asked it to guess if it would rain, it would be right 93% of the time. If you asked it to guess the intensity, it matched the real weather patterns almost perfectly.
The Twist: Unlike other models that just guess based on patterns, this one was designed to understand why things happen. The researchers checked its "brain" and confirmed it learned the real physics: trees affect the air, and the air affects the rain.

2. The Experiment: Simulating a "What If" Forest

To see what happens when trees disappear, the researchers didn't wait years for real deforestation. Instead, they used the AI to run "counterfactual" experiments. They told the model: "Imagine we cut down 1% of the trees in a specific area today. What happens to the rain tomorrow, next week, and next month?"

They tested different scenarios, from slow, steady cutting to rapid, heavy deforestation.

3. The Big Surprise: The Rain Doesn't Just Stop; It Changes Shape

The most important finding is that losing trees doesn't just make it rain less overall; it reshapes the rain in a weird way.

The "Heavy Rain" Vanishes: The model predicts that heavy, soaking rains (the kind that fill rivers and recharge groundwater) drop significantly—by up to 7%.
The "Drizzle" Increases: At the same time, light, misty drizzles actually increase by nearly 4%.
The Analogy: Imagine a healthy forest rain system is like a powerful, steady waterfall. When you cut down the trees, the waterfall doesn't just get smaller; it turns into a weak, misty spray. You get more "mist" (light rain) but much less "waterfall" (heavy rain). This is bad for the ecosystem because plants and rivers need the heavy rain, not just the mist.

4. The "Tipping Point": The 2-to-3 Month Warning

The study found a scary "tipping point" behavior.

The Slow Burn: If you start cutting trees, the rain doesn't change immediately. The system seems to hold on for a while.
The Snap: However, if the cutting continues in the most sensitive areas (like the north-western Amazon and the foothills of the Andes) for about 2 to 3 months, the system suddenly snaps. The area covered by rain shrinks dramatically and abruptly.
The Metaphor: Think of it like a rubber band. You can pull it for a while, and it stretches slowly. But once you pull it past a certain point, it doesn't just stretch more; it snaps. The model suggests the Amazon's rain system has a similar "snap" point that happens relatively quickly (in months, not decades) if the stress is sustained.

5. The "One-Way Street" Problem

The researchers also tested what happens if we plant trees back (reforestation).

The Result: Planting trees helps, but it doesn't simply "undo" the damage in reverse. The system shows hysteresis (a fancy word for a "lag" or "memory").
The Analogy: It's like a broken bone. If you break your leg, it takes a long time to heal, and even after it heals, it might not be exactly as strong as it was before. Similarly, the model suggests that even if we stop cutting trees and start planting them, the rain system might not bounce back to its original, healthy state immediately or completely. The damage lingers.

6. Where is the Danger Zone?

The model pinpointed specific "hotspots" where the rain is most fragile.

The Danger Zones: The north-western part of the Amazon and the areas right at the base of the Andes mountains are the most sensitive.
The Buffer: Interestingly, the deep interior of the Amazon and areas near big rivers seem to act as a "buffer," holding up better than the edges.

Summary

In short, this paper uses a highly accurate AI to show that the Amazon's rain system is more fragile and faster to react than we thought.

Cutting trees changes the type of rain: Less heavy rain, more light drizzle.
It happens fast: A sustained loss of trees for just 2–3 months in sensitive areas can cause a sudden, sharp drop in rainfall coverage.
It's hard to fix: Once the system is disturbed, simply planting trees back doesn't instantly fix the problem.

The study acts as an early-warning system, suggesting that we don't need to wait decades to see the effects of deforestation; the rain system can reorganize and break down on a timescale of months.

Technical Summary: Interpretable Rainfall Modelling Reveals Rapid Reorganisation of Amazonian Rainfall Under Vegetation Loss

Problem Statement
Understanding how vegetation loss alters rainfall in the Amazon remains a critical challenge due to the heterogeneous, seasonal, and nonlinear nature of land-atmosphere feedbacks. Existing modelling systems face significant limitations: convection is often parameterised at coarse spatial scales, potential tipping behaviors are poorly constrained, and analyses are frequently limited to multi-decadal timescales. Consequently, many approaches resolve correlations rather than causal effects, hindering the ability to anticipate hydrological disruption on short timescales relevant to water-security planning. Furthermore, traditional numerical weather prediction and climate models often struggle to accurately represent convective precipitation and lack mechanisms for attribution beyond correlation, particularly regarding the rapid, localised dynamics of land-atmosphere coupling.

Methodology
The authors developed an interpretable, data-driven modelling pipeline using a MultiTask Convolutional Long Short-Term Memory (ConvLSTM) network. The model was trained on ERA5-Land hourly reanalysis data (2021–2024) covering the central Amazon Basin (81 × 97 grid, ~0.25° resolution).

Architecture & Inputs: The model jointly predicts hourly rainfall occurrence (classification) and intensity (regression). Inputs include atmospheric mediators (surface fluxes, thermodynamics), precipitation memory from the previous hour, and vegetation/land-surface variables (Leaf Area Index [LAI], albedo, surface roughness, soil water content).
Training Strategy: To address the highly imbalanced distribution of precipitation data (43.65% zero-precipitation), the model employs a compound loss function combining Mean Squared Error (MSE) for regression and binary cross-entropy for classification.
Interpretability & Attribution: To ensure the model learns physically consistent pathways rather than spurious correlations, the authors employed:
- Mechanistic Ablation: Systematically removing vegetation (V), precipitation memory (P), or atmospheric (A) inputs to test dependency structures.
- Causal Graph Discovery: Using the PC algorithm to infer conditional independencies and validate indirect pathways (V → A → P).
- Gradient-Based Sensitivity: Analyzing gradients via Singular Value Decomposition (SVD) and Projected Gradient Descent (PGD) to identify dominant vegetation-rainfall sensitivity modes and minimum perturbation thresholds.
Counterfactual Experiments: The trained model was subjected to five scenarios (ranging from heavy diffuse deforestation to mild reforestation) to simulate vegetation perturbations and analyze resulting shifts in rainfall distributions, entropy, and dry-season intensity.

Key Contributions

High-Fidelity Predictive Model: Development of a MultiTask ConvLSTM that achieves strong performance in predicting hourly rainfall across the Amazon (Spearman $\rho \approx 0.84$ , F1 $\approx 0.93$ , ROC-AUC $\approx 0.98$ ), effectively capturing both dry-to-wet transitions and extreme rainfall events.
Internalisation of Mechanistic Pathways: Demonstration that the model learns temporally ordered, physically consistent dependencies where vegetation influences rainfall indirectly via atmospheric mediators (fluxes and thermodynamics), aligning with ecohydrological theory rather than relying on shortcut correlations.
Quantification of Nonlinear & Asymmetric Responses: Identification of asymmetric rainfall responses to vegetation loss versus regrowth. Heavy deforestation suppresses extreme rainfall while promoting light rainfall, whereas reforestation disproportionately enhances extreme rainfall, suggesting short-timescale hysteresis.
Threshold Behaviour Detection: Observation of a "tipping-like" behaviour where, after approximately 2–3 months of sustained vegetation perturbation in sensitive regions, the precipitating area fraction (PAF) undergoes an abrupt decline.
Spatial Heterogeneity Mapping: Identification of the north-western Amazon and Andean foothills as hotspots of short-term hydrological vulnerability, where rainfall sensitivity is most pronounced.

Key Results

Rainfall Redistribution: Under sustained deforestation (scenarios simulating 1.0–1.5% annual forest loss), heavy rainfall (20–50 mm h⁻¹) declines by up to 7%, while light rainfall (0.1–1 mm h⁻¹) increases by nearly 4%. This indicates a structural reorganisation of rainfall intensity regimes.
Dry Season Intensification: Deforestation scenarios lead to a marked increase in dry-season intensity (up to 6600 mm cumulative deficit in severe scenarios), while reforestation yields a negligible reduction in dry-season intensity, highlighting the difficulty of rapid recovery.
Entropy and Stability: Rainfall entropy increases under deforestation, indicating greater disorder in local rainfall dynamics. Interestingly, reforestation scenarios also yielded high entropy increases, suggesting that vegetation recovery may temporarily destabilise rainfall distributions due to delayed feedbacks.
Sensitivity Hierarchy: Gradient analysis reveals that evaporation is the dominant short-timescale control on rainfall (explaining 84% of variance in leading modes), followed by soil moisture, surface roughness, albedo, and LAI.
Rapid Reorganisation: The precipitating area fraction remains stable during the initial months of vegetation perturbation but declines sharply after 2–3 months in the most sensitive regions, indicating a rapid, nonlinear response to sustained stress.

Significance and Claims
The paper claims that data-driven methods can successfully uncover process-relevant signatures of land-atmosphere coupling at convective timescales (hourly), offering a complementary perspective to process-based climate models. By resolving rainfall responses at high temporal resolution, the study identifies early precursors to long-term shifts that are often obscured by monthly or seasonal averages.

The authors emphasise that their findings demonstrate the urgency of Amazon conservation, revealing that rainfall organisation can exhibit tipping-like responses to sustained vegetation degradation on timescales of months rather than decades. They argue that these results support a shift in policy from aggregate deforestation metrics toward short-term, spatially targeted assessments of rainfall sensitivity to specific land-surface characteristics.

Limitations and Modesty
The authors explicitly note several limitations:

Temporal Scope: Training on only 1.5 years of data limits exposure to interannual variability and long-timescale feedbacks (e.g., El Niño).
Causal Inference: While the model internalises structures consistent with physical mechanisms, the study does not claim to satisfy formal causal identification conditions (e.g., the front-door criterion) due to the potential for hidden confounders and the underdetermined nature of Earth system modelling. The causal analyses are presented as hypothesis-generating rather than definitive policy conclusions.
Generalisation: The model's performance degrades for intermediate rainfall intensities (2–5 mm h⁻¹) where convective processes are fragmented, and the fixed grid limits the inclusion of external teleconnections.
Counterfactual Validity: The counterfactual experiments simulate hypotheses within the model's learned dynamical regime and do not guarantee physical correctness, particularly for reforestation scenarios which may lie outside the distribution of the training data (dominated by forest loss).

In summary, the paper positions its interpretable neural framework as a valuable tool for generating ecohydrological hypotheses and identifying short-term vulnerability pathways, rather than as a replacement for long-term physical climate modelling.