🤖 machine learning

Deep Clustering for Climate: Analyzing Teleconnections through Learned Categorical States

This paper demonstrates that using Masked Siamese Networks to perform self-supervised deep clustering on temperature time series can effectively discretize complex climate data into semantically meaningful states that reflect real-world phenomena like El Niño events.

Original authors: Lívia Meinhardt, Dário Oliveira

Published 2026-04-28

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Lívia Meinhardt, Dário Oliveira

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

The "Climate Mood Ring" Approach: Making Sense of Weather Chaos

Imagine you are trying to describe the "vibe" of a massive, crowded music festival. You could try to record the exact decibel level of every single person, the precise temperature of every square inch of the field, and the exact movement of every person at every second. That would be a mountain of data—too much for any human to make sense of.

Instead, you might just say: "It’s a high-energy, hot, midday vibe," or "It’s a chill, breezy, late-night vibe."

That is exactly what this research paper is doing with the climate.

The Problem: Too Much Noise

The Earth’s weather is incredibly messy. If you look at raw temperature data from the Brazilian Cerrado (a massive tropical savanna), it looks like a chaotic sea of numbers. It’s hard to see the "big picture" because there is so much tiny, daily noise. Scientists want to know how big global events—like El Niño (a massive warming of the Pacific Ocean)—change the "mood" of the local weather in Brazil. But how do you track a "mood" using only raw numbers?

The Solution: The AI "Mood Ring"

The researchers used a specialized type of Artificial Intelligence called a Masked Siamese Network. Think of this AI as a master pattern-recognizer.

Here is how it works using a metaphor:
Imagine I show you a picture of a face, but I put a piece of tape over the eyes. Even with the eyes covered, you can probably tell if the person is smiling or frowning based on their mouth and cheeks. The AI does this with weather maps. It "hides" parts of the weather data and then tries to guess what the hidden parts look like. By doing this millions of times, it learns the "essence" of what a specific weather pattern looks like.

Instead of seeing a million different temperatures, the AI groups them into 30 "Climate Moods" (or Regimes).

Mood 1 might be: "A typical, mild summer afternoon."
Mood 2 might be: "A dangerous, extreme heatwave."
Mood 22 might be: "A sticky, warm night where the temperature never drops."

By turning messy numbers into these 30 clear "moods," the scientists turned a chaotic ocean of data into a simple, readable story.

The Discovery: How El Niño Changes the Story

Once the AI created these "moods," the researchers looked at how El Niño acts like a conductor changing the music. They found three fascinating things:

The "Heat Switch": When El Niño happens, it doesn't just make everything "a little warmer." It specifically flips the switch to certain "Extreme Heat" moods (like Mood 2 and Mood 22) while making other moods disappear.
The "Echo Effect" (Lagged Response): El Niño doesn't hit all at once. It’s like a wave hitting a shore. The researchers found that some weather moods appear before El Niño peaks, some happen exactly when it peaks, and some show up as an "aftershock" months later.
The "Changing Climate" Warning: Perhaps most importantly, they noticed that these extreme heat moods are becoming more frequent and more intense in recent decades. The "mood swings" of the Brazilian climate are getting more dramatic over time.

Why Does This Matter?

By teaching AI to simplify the weather into these "categorical states," we get a much better toolkit for the future.

Instead of just saying "it might be hot," farmers, city planners, and disaster relief teams can say: "We are entering a period where the 'Extreme Heat' mood is becoming the dominant pattern." It turns overwhelming data into actionable intelligence, helping us prepare for a world where the climate's "moods" are becoming increasingly unpredictable.

Technical Summary: Deep Clustering for Climate

Title: Deep Clustering for Climate: Analyzing Teleconnections through Learned Categorical States
Authors: Lívia Meinhardt and Dário Oliveira (FGV EMAp)

1. Problem Statement

Climate variability is characterized by high-dimensional, noisy, and non-linear dependencies. Traditional methods for identifying climate regimes often rely on manual feature engineering or scalar indices, which may fail to capture the full complexity of spatial-temporal patterns. There is a need for a method that can automatically discretize continuous, high-dimensional climate data (like temperature grids) into semantically meaningful, low-dimensional "regimes" that can be used for downstream scientific analysis, such as studying teleconnections (e.g., the influence of El Niño on regional weather).

2. Methodology

The authors propose a self-supervised framework that treats discretization as a joint representation learning and clustering problem.

Architecture: They utilize a Vision Transformer (ViT-Small) backbone.
Learning Framework: The model employs Masked Siamese Networks (MSN).
- Multi-View Generation: The input (daily $T_{min}$ and $T_{max}$ grids) undergoes Random Resized Cropping. A "target view" (large crop) is compared against "anchor views" (smaller crops with 15% masking).
- Self-Supervised Objective: The model learns by matching the latent representations of these views. It uses an anchor encoder (trainable) and a target encoder (Exponential Moving Average of the anchor weights).
Discretization (Clustering): Instead of a standard classification head, the latent space is discretized using $K=30$ learnable prototypes.
- The probability of a representation belonging to a cluster is calculated via a softmax function with a temperature parameter $\tau$ .
- Regularization: To prevent "cluster collapse" (where the model only uses a few clusters), they implement Mean Entropy Maximization (ME-MAX), ensuring all $K$ clusters are utilized.
Inference: At test time, the unmasked input is passed through the target encoder, and the resulting representation is assigned to the prototype with the highest cosine similarity, creating a deterministic categorical sequence of climate states.

3. Key Contributions

Self-Supervised Discretization: Demonstrates that a ViT trained via MSN can automatically discover physically consistent climate regimes without labeled data.
Multi-Scale Thermal Characterization: Shows that the learned clusters capture not just mean temperature shifts, but also distinct parts of the temperature distribution (extremes vs. averages).
Teleconnection Framework: Provides a novel way to analyze the relationship between large-scale climate drivers (ENSO) and regional weather by observing how the probability distribution of learned regimes shifts over time.

4. Results

The model was evaluated on Brazilian Cerrado temperature data (1961–2024).

Interpretable Regimes: The 30 learned clusters are both seasonally structured (forming meta-clusters for Summer, Winter, etc.) and thermally distinct. They successfully separate warm/cold extremes and transitional states.
ENSO (El Niño) Analysis:
- Sensitivity: The model identified specific regimes (notably Clusters 2 and 22) that are highly sensitive to El Niño. Cluster 2 is linked to intense heat extremes ( $T_{max}$ ), while Cluster 22 is linked to persistent warming ( $T_{min}$ ).
- Non-Stationarity: The sensitivity to ENSO has intensified in recent decades, particularly in the last 30 years.
- Temporal Dynamics (Lags): The analysis revealed that ENSO does not affect all regimes simultaneously. Different clusters show precursor, synchronous, or delayed responses relative to the Oceanic Niño Index (ONI), suggesting a structured "reallocation" of climate states throughout the ENSO lifecycle.
- Seasonal Modulation: The ENSO response is "filtered" by the local season; the impact of El Niño on regime frequency depends heavily on the calendar month.

5. Significance

This work bridges the gap between deep learning and climate science. By converting continuous, high-dimensional grids into a compact "alphabet" of climate states, the authors provide a tool that:

Simplifies Complexity: Reduces massive datasets into manageable categorical sequences for probabilistic modeling.
Enhances Interpretability: Moves away from "black-box" AI by producing clusters that correspond to recognizable physical phenomena.
Enables Advanced Analysis: Facilitates the study of non-stationary teleconnections and the evolution of extreme weather patterns under changing climate conditions.

Future Directions: The authors suggest expanding the model to include precipitation and humidity, and applying it to other climate modes like the Madden–Julian Oscillation (MJO).