Exploring Using AI for Early Detection of Climate Change Signals

Sungduk Yu is an AI research scientist at Intel Labs specializing in multimodal AI systems and their applications in climate science. Co-authors Anahita Bhiwandiwalla and Yaniv Gurwicz are AI research scientists focused on the intersection of multimodal AI models and hardware-aware efficient methodologies, and causal machine learning with a focus on applications in climate science respectively.

Highlights

• ClimDetect, a benchmark dataset of over 816,000 climate data samples, enables the standardization of climate change detection using artificial intelligence (AI) , promoting the development and application of AI methods for addressing climate change.
• This novel application of vision transformers (ViTs) demonstrates advanced capabilities in early detection of climate change signals, unlocking new opportunities for AI in environmental monitoring.
• ViTs can recover "fingerprints" through interpretable methods, providing physical grounding for AI-based insights.
  
  

Climate change is among the most critical environmental challenges of our time. As the Intergovernmental Panel on Climate Change (IPCC) has emphasized, accurately identifying human-driven influences is critical for guiding effective mitigation and adaptation strategies. However, detecting these anthropogenic signals, or changes in the environment caused by human activity, is particularly challenging. Natural climate variability can obscure these indicators on similar spatial and temporal scales. To improve model accuracy in identifying signals, Intel Labs introduced ClimDetect, an open source benchmark dataset with more than 816,000 climate data points for detecting climate change signals in daily weather patterns. Using vision transformers for image tasks, this novel approach shows great promise in analyzing spatial climate data.

Early detection of climate change signals and identifying climate change fingerprints — characteristic spatial patterns of climate change — can provide tangible benefits. By recognizing warming patterns in specific regions, communities can take timely actions such as protecting ecosystems, adapting agricultural practices, or strengthening infrastructure to handle extreme weather events. Proactive planning based on early signals can minimize impacts and improve resource allocation for climate adaptation.

This research reflects Intel's commitment to responsible AI, emphasizing transparency, accountability, and environmental protection. By making ClimDetect openly available, we aim to inspire further innovation in climate science and demonstrate how AI can make a positive impact on global environmental challenges. The dataset is preprocessed and training ready, enabling both climate scientists and AI researchers to contribute solutions.

The Challenge of Climate Change Signal Detection

Natural variability, driven by the Earth's complex climate system, produces fluctuations that can last weeks to decades, often overshadowing the slowly evolving anthropogenic signal. This creates what scientists describe as a "signal in noise" problem, where the human-caused signals are masked by natural patterns. This difficulty is especially pronounced in the early stages of global warming, where the signal is faint and deeply intertwined with natural variability. However, early detection is crucial for enabling resource-efficient climate adaptation actions, as it allows for timely interventions that mitigate risks and optimize resource allocation.

Addressing a Key Gap with AI

Traditional detection methods rely on linear statistical models and require long-term historical climate records. The need for lengthy observation records restricts traditional methods in handling the nonlinearity inherent in complex climate systems and the sensitivity required for early detection of signals. Hence, we explored how AI technologies, with their ability to handle complex nonlinear systems and process vast datasets, can overcome these limitations. In particular, we focused on how the explicit global attention mechanisms in vision transformers can enable earlier and more sensitive detection of anthropogenic signals.

We trained popular ViT models using global climate model outputs from the Coupled Model Intercomparison Project Phase 6 (CMIP6). We tested these models against reanalysis products, such as the European Centre for Medium-Range Weather Forecasts (ECMWF) ERA5, which combines observational data with modeling to provide comprehensive and physically consistent climate states. Alongside ViTs, we trained simpler models such as ridge regression, multi-layer perceptron (MLP), and convolutional neural network (CNN) as a benchmark. The models were tasked with using three daily climate global snapshots — surface air temperature, humidity, and precipitation — to predict annual global mean temperature (AGMT), a popular metric to encapsulate the state of climate change. To push the limits further, we conducted a more challenging "mean-removed" experiment by removing global mean values from the input data, forcing models to rely solely on spatial patterns.

We then evaluated model performance using the year of emergence (YOE), defined as the first year when climate change signals statistically surpass natural variability. An early YOE indicates a model's superior sensitivity to climate signals.

Figure 1. This figure illustrates our climate change signal detection methodology. The diagram shows climate field maps, color-coded to represent different datasets: The training dataset (orange), the historical dataset (green, pre-warming), and the observation (purple). Fθ represents a detection model, such as a vision transformer, where θ refers to the model’s parameters. Each purple dot denotes an estimate derived from a single observation sample.

A Breakthrough in Detection Capabilities with AI

ViTs detect the climate change signal earlier than the simpler models, as shown in Figure 2. Even under the challenging mean-removed setup, ViTs performance advantage becomes more apparent, demonstrating their ability to extract meaningful spatial anomalies effectively. These results underscore ViTs' potential as advanced tools for climate signal detection.

Figure 2. Bar chart, on left: Year of emergence represents the first year when the majority of daily climate fields show a detectable climate change signal beyond natural variability. Gray bars indicate cases where models failed to capture YOE within the reanalysis period (1980-2023). Global maps, on center and right: Integrated gradients visualization from the mean-removed experiment highlights regions influencing AGMT predictions. Values are normalized by the maximum value, where -1 corresponds to the coldest color and +1 to the warmest color.

Physical interpretability remains crucial for establishing data-driven models as a new tool in climate science and is equally vital for responsible AI use, especially for the transparency needed in decision-making. We show preliminary model interpretations using integrated gradients for the mean-removed experiment, revealing differences between nonlinear AI models and ridge regression. Unlike ridge regression, which focuses on land-sea contrasts, ViTs show a greater positive dependence on the Antarctic Ocean, offering physically meaningful insights. These patterns provide valuable perspectives on the potential hotspots for climate change detection, which was hidden in a linear framework.

What's Next?

This work represents just one example of Intel's efforts to leverage state-of-the-art AI technologies for environmental protection. Looking ahead, we are exploring multimodal vision-language models and agentic AI frameworks to provide expert-level climate insights easily accessible to users. Stay tuned for more advancements as we continue to harness AI to address critical environmental challenges.