Search

Artificial intelligence (AI) and machine learning (ML) pose a challenge for achieving science that is both reproducible and replicable. The challenge is compounded in supervised models that depend on manually labeled training data, as they introduce additional decision‐making and processes that require thorough documentation and reporting. We address these limitations by providing an approach to hand labeling training data for supervised ML that integrates quantitative content analysis (QCA)—a method from social science research. The QCA approach provides a rigorous and well‐documented hand labeling procedure to improve the replicability and reproducibility of supervised ML applications in Earth systems science (ESS), as well as the ability to evaluate them. Specifically, the approach requires (a) the articulation and documentation of the exact decision‐making process used for assigning hand labels in a "codebook" and (b) an empirical evaluation of the reliability" of the hand labelers. In this paper, we outline the contributions of QCA to the field, along with an overview of the general approach. We then provide a case study to further demonstrate how this framework has and can be applied when developing supervised ML models for applications in ESS. With this approach, we provide an actionable path forward for addressing ethical considerations and goals outlined by recent AGU work on ML ethics in ESS. Plain Language Summary: Artificial intelligence and machine learning can make it hard to do science in a way that can be repeated. This can mean redoing a study in the exact same way to see if you can get the same or similar results (reproducibility) or trying to use the same study design on a new problem to see if the results are the same or similar (replicability). These types of scientific repetitions is important for developing robust knowledge, but is hard to do with certain types of machine learning that rely on data that were categorized by researchers. The researchers have to make decisions and categorize their data, which the machine learning algorithm then uses as a guide to make its own decisions. Generally, there is not enough information shared by the researchers about how these decisions were made to repeat the science or evaluate how good it is. In this paper, we provide a way to address these shortcomings. The approach and example we offer illustrates how to (a) create a rulebook that can be shared for how to make decisions and (b) quantitatively measure how consistent the researchers are at using that rulebook to make their decisions. Key Points: We provide a rigorous hand labeling procedure to improve the replicability and reproducibility of supervised machine learning (ML)Our case study and step‐by‐step guide clearly outline how the procedure can be appliedThe procedure is an actionable path forward for addressing ethical considerations and goals for ML development in Earth systems science [ABSTRACT FROM AUTHOR]

Earthquake nowcasting has been proposed as a means of tracking the change in large earthquake potential in a seismically active area. The method was developed using observable seismic data, in which probabilities of future large earthquakes can be computed using Receiver Operating Characteristic methods. Furthermore, analysis of the Shannon information content of the earthquake catalogs has been used to show that there is information contained in the catalogs, and that it can vary in time. So an important question remains, where does the information originate? In this paper, we examine this question using stochastic simulations of earthquake catalogs. Our catalog simulations are computed using an Earthquake Rescaled Aftershock Seismicity ("ERAS") stochastic model. This model is similar in many ways to other stochastic seismicity simulations, but has the advantage that the model has only 2 free parameters to be set, one for the aftershock (Omori‐Utsu) time decay, and one for the aftershock spatial migration away from the epicenter. Generating a simulation catalog and fitting the two parameters to the observed catalog such as California takes only a few minutes of wall clock time. While clustering can arise from random, Poisson statistics, we show that significant information in the simulation catalogs arises from the "non‐Poisson" power‐law aftershock clustering, implying that the practice of de‐clustering observed catalogs may remove information that would otherwise be useful in forecasting and nowcasting. We also show that the nowcasting method provides similar results with the ERAS model as it does with observed seismicity. Plain Language Summary: Earthquake nowcasting was proposed as a means of tracking the change in the potential for large earthquakes in a seismically active area, using the record of small earthquakes. The method was developed using observed seismic data, in which probabilities of future large earthquakes can be computed using machine learning methods that were originally developed with the advent of radar in the 1940s. These methods are now being used in the development of machine learning and artificial intelligence models in a variety of applications. In recent times, methods to simulate earthquakes using the observed statistical laws of earthquake seismicity have been developed. One of the advantages of these stochastic models is that it can be used to analyze the various assumptions that are inherent in the analysis of seismic catalogs of earthquakes. In this paper, we analyze the importance of the space‐time clustering that is often observed in earthquake seismicity. We find that the clustering is the origin of information that makes the earthquake nowcasting methods possible. We also find that a common practice of "aftershock de‐clustering", often used in the analysis of these catalogs, removes information about future large earthquakes. Key Points: Earthquake nowcasting tracks the change in the potential for large earthquakes, using information contained in seismic catalogsWe analyze the information contained in the space‐time clustering that is observed in earthquake seismicityWe find that "aftershock de‐clustering" of catalogs removes information about future large earthquakes that the nowcasting method uses [ABSTRACT FROM AUTHOR]

Accurate rainfall measurement with a precise spatial and temporal resolution is essential for making informed decisions during disasters and conducting scientific studies, particularly in regions characterized by intricate terrain and limited coverage of automated weather stations. Retrieval of precipitation with satellite is currently the most effective means to obtain precipitation over large‐scale areas. The key to enhancing the accuracy of precipitation estimation and forecasting in regions with complex terrain lies in effectively integrating satellite data with topographic information. This paper introduces a deep learning approach called AttUnet_R_SFT that utilizes high temporal, spatial, and spectral resolution data obtained from the Fengyun 4A satellite, and incorporates the Deep Spatial Feature Transform (SFT) layer to incorporate geographical data for estimating half‐hourly precipitation in northeastern China. We assess it by compared to operational near‐real‐time satellite precipitation products demonstrated to be successful in estimating precipitation and baseline deep learning models. According to the experimental findings, the AttUnet_R_SFT model outperforms practical precipitation products and baseline deep learning models in both identifying and estimating precipitation. The main enhancement of the model performance is shown in the windward slope of the Greater Khingan Mountains as a result of the successful incorporation of geographical data. Hence, the suggested framework holds the capability to function as a superior and dependable satellite‐derived precipitation estimation solution in regions characterized by intricate terrain and infrequent rainfall. The findings of this study indicate that the utilization of deep learning algorithms for satellite precipitation estimation shows potential as a fruitful avenue for further research. Plain Language Summary: A deep learning model named AttUnet_R_SFT is proposed, which use high temporal, spatial and spectral resolution data from the Fengyun 4A satellite, and combines with the Deep Spatial Feature Transform (SFT) layer to input geographic information for half‐hourly precipitation estimation in the complex terrain region represented by northeast China. The model can provide a reference for improving the performance of precipitation estimation in areas with complex topography. Key Points: A deep‐learning model is proposed to effectively fuse satellite multispectral data of the Fengyun 4A satellite, with topographic informationHalf‐hourly precipitation is estimated with higher temporal resolution, which is closer to the operational needs of weather forecastingAs precipitation in the study area is a non‐high‐frequency event, data enhancement is attempted to use and obtain effective results [ABSTRACT FROM AUTHOR]

Inversion of magnetic basement interfaces in basins is essential for interpreting potential field data and studying geothermal resource distribution, as well as basin formation and evolution. This paper introduces a novel method for inverting magnetic basement interfaces using a random forest regression (RFR) algorithm that combines potential field processing and machine learning techniques. The method creates magnetic base interface models and corresponding magnetic anomaly data through the random midpoint displacement method and magnetic interface finite element forward simulation. These anomalies are then processed using techniques such as directional transformations, analytical continuation, spatial derivatives, and fractional transformations. Feature attributes are extracted, and Gini importance is utilized to measure the contributions of feature factors, identify effective features, and enhance algorithm efficiency. The validity and practicality of the method are demonstrated through the analysis of both idealized and noisy models. The proposed machine learning‐based approach is more intelligent, efficient, and accurately represents the relief of magnetic base interfaces. When applied to magnetic survey data in the Datong Basin, it produced a reliable basin base model that aligns with known structural information, paving the way for further research in magnetic interface inversion. Key Points: A machine learning method based on random forest regression algorithm is used to recover the relief of the magnetic basement interface by using high‐precision magnetic dataApplying potential field processing and transformation to feature extraction for machine learningThe proposed technique intelligently and efficiently restores the 3D relief of the magnetic basement interface, and it has been successfully implemented in the Datong Basin's Tianzhen‐Yanggao sag [ABSTRACT FROM AUTHOR]

Tsunami simulations require powerful computational resources to be performed efficiently. Although the modern graphics processing units (GPUs) allow the acceleration of this kind of simulations, they can still last many minutes or even hours for simulations which have to deal with very high spatial resolutions or simulation times. In this paper, we propose a method to predict the alert or inundation level of a tsunami generated by an earthquake using deep learning methods. In particular, we train multilayer perceptron (MLP) neural networks for predicting the alert level due to a tsunami at given coastal locations. Ensemble methods are used to improve the predictions of the neural networks. Tsunamis caused by ruptures of several fault segments at different time instants, application to real events, probabilistic forecasting and comparison with other machine learning algorithms are also addressed. Results on realistic scenarios confirm that good accuracies are obtained. The inference times of the trained networks and ensembles are also very low, lasting less than one second to predict the results of thousands of simulations. The proposed method could be used in a tsunami early warning system along with the application of scaling laws. Plain Language Summary: In tsunami early warning systems (TEWS) it is crucial to perform fast predictions of the risk of a tsunami in coastal areas in order to take appropriate measures before the arrival of the tsunami. Currently, tsunami simulations may take many minutes or even hours despite using modern computational resources. In this work we propose a method based on deep learning techniques to predict the alert level of a tsunami generated by an earthquake at given coastal locations in a fraction of a second. This method could be used in TEWS along with techniques to estimate the values of some parameters of the earthquake. Application to real events and probabilistic predictions are also addressed. Key Points: A deep learning method to predict different tsunami alert levels is proposedTsunamis caused by the rupture of multiple fault segments at different time instants are supportedApplication to real events with probabilistic predictions are addressed [ABSTRACT FROM AUTHOR]

Satellite infrared detectors cannot penetrate clouds, especially precipitating clouds. Improving precipitation estimation accuracy based on infrared brightness temperature has always been important but challenging. In this paper, based on the infrared brightness temperature of the Advanced Geosynchronous Radiation Imager (AGRI) onboard China's Feng‐Yun 4A satellite, we develop and evaluate a new precipitation estimation method. First, using static data, physical characteristics of clouds, cloud image texture features, temporal motion features, and AGRI infrared channel brightness temperature, we construct features for a machine learning model. Then, we develop precipitation estimation methods. Precipitation is estimated in two steps: classification and regression. We employ a random forest classification model to identify whether there is precipitation in a given field of view. If there is precipitation, a multi‐model ensemble regression learning method is used to estimate the areas with this precipitation. The ensemble learning method uses convex optimization to integrate prediction results based on the optimization of hyperparameters of five basic models (i.e., those of random forest, XGBoost, LightGBM, decision tree, and extra tree models). Furthermore, two regression stacking ensemble models—the Least Absolute Shrinkage and Selection Operator (herein referred to as Stacking1‐LASSO) and K‐nearest neighbor (herein referred to as Stacking2‐KNN)—are used to predict the results of the aforementioned basic models. The results of basic models are used as inputs of these two stacking models. Finally, based on the Integrated Multi‐satellitE Retrievals for GPM (IMERG) precipitation product and rain gauge precipitation data, we conduct precipitation estimation experiments and evaluate our methods. The results show that ensemble learning models have greater accuracy in estimating precipitation than the basic models. When using IMERG precipitation as the target precipitation, ensemble learning models can estimate the central area of heavy precipitation during typhoons Ampil and Maria. The ensemble learning estimation effect is better than that of Stacking2‐KNN. Moreover, when rain gauge data is used as the target precipitation, ensemble learning can also estimate the center of heavy precipitation and with good consistency with recorded satellite brightness temperature data. Key Points: A convex optimization weighted and stacked ensemble learning method for satellite data estimation of precipitation is developedNear real‐time precipitation estimates can be obtained throughout the day using only geographic information and Advanced Geosynchronous Radiation Imager infrared brightness temperature dataThe proposed ensemble learning method can estimate the central areas of typhoons and short‐term heavy rainfall [ABSTRACT FROM AUTHOR]

Economic growth, air pollution, and forest fires in some states in the United States have increased the concentration of particulate matter with a diameter less than or equal to 2.5 μm (PM2.5). Although previous studies have tried to observe PM2.5 both spatially and temporally using aerosol remote sensing and geostatistical estimation, they were limited in accuracy by coarse resolution. In this paper, the performance of machine learning models on predicting PM2.5 is assessed with linear regression (LR), decision tree (DT), gradient boosting regression (GBR), AdaBoost regression (ABR), XGBoost (XGB), k‐nearest neighbors (K‐NN), long short‐term memory (LSTM), random forest (RF), and support vector machine (SVM) using PM2.5 station data from 2017 to 2021. To compare the accuracy of all the nine machine learning models, the coefficient of determination (R2), root mean square error (RMSE), Nash‐Sutcliffe efficiency (NSE), root mean square error ratio (RSR), and percent bias (PBIAS) were evaluated. Among all nine models, the RF (100 decision trees with a max depth of 20) and support vector regression (SVR; nonlinear kernel, degree 3 polynomial) models were the best for predicting PM2.5 concentrations. Additionally, comparison of the PM2.5 performance metrics displayed that the models had better predictive behavior in the western United States than that in the eastern United States. Key Points: We present the prediction of PM2.5 concentrations over the United States using various machine learning (ML) algorithmsWe show ML as a new approach for analyzing large data sets due to the computational speed and easy implementation for massive amounts of dataThe study is important for improving our understanding of the differences among ML algorithms for Earth Science research [ABSTRACT FROM AUTHOR]

Nowcasting is a term originating from economics, finance, and meteorology. It refers to the process of determining the uncertain state of the economy, markets or the weather at the current time by indirect means. In this paper, we describe a simple two‐parameter data analysis that reveals hidden order in otherwise seemingly chaotic earthquake seismicity. One of these parameters relates to a mechanism of seismic quiescence arising from the physics of strain‐hardening of the crust prior to major events. We observe an earthquake cycle associated with major earthquakes in California, similar to what has long been postulated. An estimate of the earthquake hazard revealed by this state variable time series can be optimized by the use of machine learning in the form of the Receiver Operating Characteristic skill score. The ROC skill is used here as a loss function in a supervised learning mode. Our analysis is conducted in the region of 5° × 5° in latitude‐longitude centered on Los Angeles, a region which we used in previous papers to build similar time series using more involved methods (Rundle & Donnellan, 2020, https://doi.org/10.1029/2020EA001097; Rundle, Donnellan et al., 2021, https://doi.org/10.1029/2021EA001757; Rundle, Stein et al., 2021, https://doi.org/10.1088/1361-6633/abf893). Here we show that not only does the state variable time series have forecast skill, the associated spatial probability densities have skill as well. In addition, use of the standard ROC and Precision (PPV) metrics allow probabilities of current earthquake hazard to be defined in a simple, straightforward, and rigorous way. Plain Language Summary: Earthquake nowcasting refers to the determination of hazard for major earthquakes at the present time, the recent past, and the near future. Nowcasting is an idea borrowed from economics, markets, and meteorology, where it has been frequently used. In this paper, we show that there is order hidden within chaotic earthquake seismicity using a very simple transformation of the data. Small earthquakes appear to transition from unstable stick‐slip events that produce seismic waves, to stable sliding where no seismic waves are produced. Our hypothesis is that this transition is due to a material phenomenon called strain‐hardening, that is frequently observed in laboratory rock mechanics experiments. The result is a state variable time series, computed over the last 51 years in California, that strongly resembles the long‐anticipated cycle of stress accumulation and release. Using supervised machine learning techniques, we can optimize the two‐parameter model. From that optimized model, we can rigorously calculate the probability of current hazard from major earthquakes. Extending these methods, we can also compute spatial hazard as well. The result is a new method for assessing earthquake hazard that may be useful for a variety of applications. Key Points: "Chaotic" seismicity contains hidden structure in the form of state variable time seriesStandard data science methods can be used to convert the time series to probabilitiesBoth temporal and spatial probabilities can be computed [ABSTRACT FROM AUTHOR]

We propose a new machine learning‐based method for nowcasting earthquakes to image the time‐dependent earthquake cycle. The result is a timeseries that may correspond to the process of stress accumulation and release. The timeseries are constructed by using principal component analysis of regional seismicity. The patterns are found as eigenvectors of the cross‐correlation matrix of a collection of seismicity timeseries in a coarse grained regional spatial grid (pattern recognition via unsupervised machine learning). The eigenvalues of this matrix represent the relative importance of the various eigenpatterns. Using the eigenvectors and eigenvalues, we compute the weighted correlation timeseries of the regional seismicity. This timeseries has the property that the weighted correlation generally decreases prior to major earthquakes in the region, and increases suddenly just after a major earthquake occurs. As in a previous paper (Rundle & Donnellan, 2020, https://doi.org/10.1029/2020ea001097), we find that this method produces a nowcasting timeseries that resembles the hypothesized regional stress accumulation and release process characterizing the earthquake cycle. We then address the problem of whether the timeseries contain information regarding future large earthquakes. For this, we compute a receiver operating characteristic and determine the decision thresholds for several future time periods of interest (optimization via supervised machine learning). We find that signals can be detected that can be used to characterize the information content of the timeseries. These signals may be useful in assessing present and near‐future seismic hazards. Plain Language Summary: Major earthquakes on fault systems in a tectonically active region are thought to occur in approximately repetitive cycles as a result of the buildup and release of tectonic forces (stress). Nowcasting is a technique adopted from weather, finance, and other fields that use readily observable proxy data to represent the unobservable stress accumulation process of interest. This paper presents a method that computes a timeseries representing the weighted correlation of small earthquake activity in the California region from 1950 to 2020. Prior to major magnitude M > 7 earthquakes, the timeseries trends toward lower values. Just after the earthquake occurs, the timeseries increases suddenly in association with the earthquake, before resuming its gradual trend toward lower values. Plotting the timeseries on an inverted scale, one sees a cyclic behavior that strongly resembles the hypothesized earthquake cycle. In principle, we can therefore use this timeseries for nowcasting, as a proxy for stress accumulation and release. Using methods of signal detection first developed for radar by the British in the 1940's, we find that the timeseries contain information about future large earthquakes that can be used for hazard assessment. Key Points: The current state of the earthquake cycle of tectonic stress accumulation and release is unobservable with existing methodsWe show that readily observable small earthquake correlations can be used to nowcast the current state of the earthquake cycleMachine learning techniques indicate that signals corresponding to future large earthquakes can be detected in a correlation time series [ABSTRACT FROM AUTHOR]

In most data assimilation and numerical weather prediction systems, the Gaussian assumption is prevalent for the behavior of the random variables/errors that are involved. At the Cooperative Institute for Research in the Atmosphere theory has been developed for different forms of variational data assimilation schemes that enables the Gaussian assumption to be relaxed. For certain variable types, a lognormally distributed random variable can be combined in a mixed Gaussian‐lognormal distribution to better capture the interactions of the errors of different distributions. However, assuming that a distribution can change in time, then developing techniques to know when to switch between different versions of the data assimilation schemes becomes very important. By dynamically changing the formulation of the data assimilation system we are able to assimilate observations in a way that reflects the flow‐dependent variability of their distribution. In this paper, we present results with a machine learning technique (the support vector machine) to switch between data assimilation methods based on the detection of a change in the Lorenz 1963 model's z component's probability distribution. Given the machine learning technique's detection/prediction of a change in the distribution, then either a Gaussian or a mixed Gaussian‐lognormal 3DVar based cost function is used to minimize the errors in this period of time. It is shown that switching from a Gaussian 3DVar to a lognormal 3DVar at lognormally distributed parts of the attractor improves the data assimilation analysis error compared to using one distribution type for the entire attractor. Key Points: Machine learning can be used to detect non‐Gaussian distributionsGiven the machine learning detection, we can switch between optimal data assimilation schemesData assimilation can be improved using machine learning techniques [ABSTRACT FROM AUTHOR]

Classifying images using supervised machine learning (ML) relies on labeled training data—classes or text descriptions, for example, associated with each image. Data‐driven models are only as good as the data used for training, and this points to the importance of high‐quality labeled data for developing a ML model that has predictive skill. Labeling data is typically a time‐consuming, manual process. Here, we investigate the process of labeling data, with a specific focus on coastal aerial imagery captured in the wake of hurricanes that affected the Atlantic and Gulf Coasts of the United States. The imagery data set is a rich observational record of storm impacts and coastal change, but the imagery requires labeling to render that information accessible. We created an online interface that served labelers a stream of images and a fixed set of questions. A total of 1,600 images were labeled by at least two or as many as seven coastal scientists. We used the resulting data set to investigate interrater agreement: the extent to which labelers labeled each image similarly. Interrater agreement scores, assessed with percent agreement and Krippendorff's alpha, are higher when the questions posed to labelers are relatively simple, when the labelers are provided with a user manual, and when images are smaller. Experiments in interrater agreement point toward the benefit of multiple labelers for understanding the uncertainty in labeling data for machine learning research. Plain Language Summary: After hurricanes and storms, pictures taken from a plane can be used to observe how the coast was impacted. A single flight might take thousands of pictures. If a computer could automatically analyze the pictures, then a person would not need to look at them one‐by‐one. To teach a computer to analyze images, we need many pictures and many labels that describe what is visible in each picture. But where do we get those labels? Typically, a coastal scientist labels the pictures by sorting them into folders or typing codes into a spreadsheet. But does every coastal scientist label pictures the same way? Some labeling questions are easy to answer, and scientists mostly agree ("Is this image all water?"). Other labeling questions are harder to answer, and cause disagreement ("Was there damage to buildings?"). This paper is about how well scientists agree when labeling the same pictures, and how we can improve agreement among scientists. We try some experiments and offer a few ideas on how to improve agreement. We suggest writing very clear questions, using smaller images, and having a comprehensive manual. It turns out that having a manual with examples—and reading the manual!—really helps. Key Points: We measure agreement among coastal scientists labeling the same sets of poststorm imagesCoastal scientists agree more when rating landforms, less when labeling inferred processesIterating on questions, providing documentation, and using smaller image sizes all increase agreement [ABSTRACT FROM AUTHOR]