Estimates of sea surface temperature (SST) variability are crucial for examining changes in the El Niño–Southern Oscillation, evaluating connections between climate variability and tropical cyclones, and pinpointing the radiative feedback mechanisms that drive climate change. To improve our understanding of SST variability, researchers rely on a range of methods, including the use of instrumental records (e.g., ship logs, buoys, satellite data), proxy data (e.g., coral, ice cores, tree rings), and climate models. Historical SST observations are constrained by biases from measurement techniques and data sparsity, hindering a comprehensive understanding of modern climate change.
We introduce a novel approach leveraging denoising diffusion probabilistic models (DDPM), a generative form of deep learning adept at recovering missing data in images, to reconstruct sparse observational SST fields. These models operate on a dual-phase mechanism: initially diffusing noise into the data (forward diffusion), then diligently extracting it to recover the original information (reverse diffusion). We first use SST climate model output to train the DDPM on SST distributions. Subsequently, we employ a conditional framework to the trained DDPM. This allows the model to interpret the available SST data as foundational data, akin to a ground truth image in image inpainting, and to produce complete SST fields.
For validation and testing, we use the root-mean-squared error metric (RMSE) on unseen, artificially-masked model output SST fields to quantify the DDPM's fidelity in matching the complete ground truth field. This evaluation further extends our ability to assess the model's scalability in inpainting varying sizes of regions and its performance on both sparse and comprehensive datasets.
- Tags
-