In climate modeling, the stratospheric ozone layer is typically only considered in a highly simplified form due to computational constraints. For climate projections, it would be of advantage to include the mutual interactions between stratospheric ozone, temperature, and atmospheric dynamics to accurately represent radiative forcing. The overarching goal of our research is to replace the ozone layer in climate models with a machine-learned neural representation of the stratospheric ozone chemistry that allows for a particularly fast, but accurate and stable simulation. We created a benchmark data set from pairs of input and output variables that we stored from simulations of the ATLAS Chemistry and Transport Model. We analyzed several variants of multilayer perceptrons suitable for physical problems to learn a neural representation of a function that predicts 24-h ozone tendencies based on input variables. We performed a comprehensive hyperparameter optimization of the multilayer perceptron using Bayesian search and Hyperband early stopping. We validated our model by replacing the full chemistry module of ATLAS and comparing computation time, accuracy, and stability. We found that our model had a computation time that was a factor of 700 faster than the full chemistry module. The accuracy of our model compares favorably to the full chemistry module within a 2-year simulation run, also outperforms a previous polynomial approach for fast ozone chemistry, and reproduces seasonality well in both hemispheres. In conclusion, the neural representation of stratospheric ozone chemistry in simulation resulted in an ozone layer that showed a high accuracy, significant speed-up, and stability in a long-term simulation.

Land cover classification (LCC) and natural disaster response (NDR) are important issues in climate change mitigation and adaptation. Existing approaches that use machine learning with Earth observation (EO) imaging data for LCC and NDR often rely on fully annotated and segmented datasets. Creating these datasets requires a large amount of effort, and a lack of suitable datasets has become an obstacle in scaling the use of machine learning for EO. In this study, we extend our prior work on Scene-to-Patch models: an alternative machine learning approach for EO that utilizes Multiple Instance Learning (MIL). As our approach only requires high-level scene labels, it enables much faster development of new datasets while still providing segmentation through patch-level predictions, ultimately increasing the accessibility of using machine learning for EO. We propose new multi-resolution MIL architectures that outperform single-resolution MIL models and non-MIL baselines on the DeepGlobe LCC and FloodNet NDR datasets. In addition, we conduct a thorough analysis of model performance and interpretability.

Surface ozone is an air pollutant that contributes to hundreds of thousands of premature deaths annually. Accurate short-term ozone forecasts may allow improved policy actions to reduce the risk to human health. However, forecasting surface ozone is a difficult problem as its concentrations are controlled by a number of physical and chemical processes that act on varying timescales. We implement a state-of-the-art transformer-based model, the temporal fusion transformer, trained on observational data from three European countries. In four-day forecasts of daily maximum 8-hour ozone (DMA8), our novel approach is highly skillful (MAE = 4.9 ppb, coefficient of determination $ {\mathrm{R}}^2=0.81 $) and generalizes well to data from 13 other European countries unseen during training (MAE = 5.0 ppb, $ {\mathrm{R}}^2=0.78 $). The model outperforms other machine learning models on our data (ridge regression, random forests, and long short-term memory networks) and compares favorably to the performance of other published deep learning architectures tested on different data. Furthermore, we illustrate that the model pays attention to physical variables known to control ozone concentrations and that the attention mechanism allows the model to use the most relevant days of past ozone concentrations to make accurate forecasts on test data. The skillful performance of the model, particularly in generalizing to unseen European countries, suggests that machine learning methods may provide a computationally cheap approach for accurate air quality forecasting across Europe.

Determining an accurate picture of ocean currents is an important societal challenge for oceanographers, aiding our understanding of the vital role currents play in regulating Earth’s climate, and in the dispersal of marine species and pollutants, including microplastics. The geodetic approach, which combines satellite observations of sea level and Earth’s gravity, offers the only means to estimate the dominant geostrophic component of these currents globally. Unfortunately, however, geodetically-determined geostrophic currents suffer from high levels of contamination in the form of geodetic noise. Conventional approaches use isotropic spatial filters to improve the signal-to-noise ratio, though this results in high levels of attenuation. Hence, the use of deep learning to improve the geodetic determination of the ocean currents is investigated. Supervised machine learning typically requires clean targets from which to learn. However, such targets do not exist in this case. Therefore, a training dataset is generated by substituting clean targets with naturally smooth climate model data and generative machine learning networks are employed to replicate geodetic noise, providing noisy input and clean target pairs. Prior knowledge of the geodetic noise is exploited to develop a more realistic training dataset. A convolutional denoising autoencoder (CDAE) is then trained on these pairs. The trained CDAE model is then applied to unseen real geodetic ocean currents. It is demonstrated that our method outperforms conventional isotropic filtering in a case study of four key regions: the Gulf Stream, the Kuroshio Current, the Agulhas Current, and the Brazil-Malvinas Confluence Zone.

Complex machine learning architectures and high-dimensional gridded input data are increasingly used to develop high-performance geoscience models, but model complexity obfuscates their decision-making strategies. Understanding the learned patterns is useful for model improvement or scientific investigation, motivating research in eXplainable artificial intelligence (XAI) methods. XAI methods often struggle to produce meaningful explanations of correlated features. Gridded geospatial data tends to have extensive autocorrelation so it is difficult to obtain meaningful explanations of geoscience models. A recommendation is to group correlated features and explain those groups. This is becoming common when using XAI to explain tabular data. Here, we demonstrate that XAI algorithms are highly sensitive to the choice of how we group raster elements. We demonstrate that reliance on a single partition scheme yields misleading explanations. We propose comparing explanations from multiple grouping schemes to extract more accurate insights from XAI. We argue that each grouping scheme probes the model in a different way so that each asks a different question of the model. By analyzing where the explanations agree and disagree, we can learn information about the scale of the learned features. FogNet, a complex three-dimensional convolutional neural network for coastal fog prediction, is used as a case study for investigating the influence of feature grouping schemes on XAI. Our results demonstrate that careful consideration of how each grouping scheme probes the model is key to extracting insights and avoiding misleading interpretations.

Feature selection is an important component of machine learning for researchers that are confronted with high dimensional data. In the field of economics, researchers are often faced with high dimensional data, particularly in the studies that aim to understand the channels through which climate change affects the welfare of countries. This work reviews the current literature that introduces various feature selection algorithms that may be useful for applications in this area of study. The article first outlines the specific problems that researchers face in understanding the effects of climate change on countries’ macroeconomic outcomes, and then provides a discussion regarding different categories of feature selection. Emphasis is placed on two main feature selection algorithms: Least Absolute Shrinkage and Selection Operator and causality-based feature selection. I demonstrate an application of feature selection to discover the optimal heatwave definition for economic outcomes, enhancing our understanding of extreme temperatures’ impact on the economy. I argue that the literature in computer science can provide useful insights in studies concerned with climate change as well as its economic outcomes.

Global warming will cause unprecedented changes to the world. Predicting events such as food insecurities in specific earth regions is a valuable way to face them with adequate policies. Existing food insecurity prediction models are based on handcrafted features such as population counts, food prices, or rainfall measurements. However, finding useful features is a challenging task, and data scarcity hinders accuracy. We leverage unsupervised pre-training of neural networks to automatically learn useful features from widely available Landsat-8 satellite images. We train neural feature extractors to predict whether pairs of images are coming from spatially close or distant regions on the assumption that close regions should have similar features. We also integrate a temporal dimension to our pre-training to capture the temporal trends of satellite images with improved accuracy. We show that with unsupervised pre-training on a large set of satellite images, neural feature extractors achieve a macro F1 of 65.4% on the Famine Early Warning Systems network dataset—a 24% improvement over handcrafted features. We further show that our pre-training method leads to better features than supervised learning and previous unsupervised pre-training techniques. We demonstrate the importance of the proposed time-aware pre-training and show that the pre-trained networks can predict food insecurity with limited availability of labeled data.