Search

Accurate crop type mapping is essential for crop growth monitoring and yield estimation. Recently, various machine learning methods have been increasingly used for crop type mapping, but they often lose their validity when directly applied to other regions and years due to differences in the distribution of source and target data, that is, domain shift. To address the problem, we developed a deep adaptation crop classification network (DACCN) based on the idea of unsupervised domain adaptation (UDA). The proposed DACCN mainly consists of two parts, a feature extractor that converts the original input into high-level representations, and a domain aligner where the discrepancy between source and target distributions is measured using multiple kernel variant of maximum mean discrepancy (MK-MMD). Four states in the United States (U.S.) Corn Belt and two provinces in northeastern China were used as study areas, where samples used for model building and accuracy evaluation were collected based on time-series Sentinel-2 imagery and reference maps in 2018 and 2019. Then, three experiments were designed to verify the transferability of DACCN across space, time, and space–time, respectively. In each experiment, the proposed DACCN was compared to deep crop classification network (DCCN), a model with a similar structure to DACCN but without the domain adaptation mechanism, and two machine learning methods, random forest (RF) and support vector machines (SVM). The experimental results showed that DACCN outperformed other models in most transfer cases with overall classification accuracies ranging from 0.835 to 0.922. The DACCN also performed better in spatially continuous mapping with its predicted crop type maps more consistent with the reference ones. As an innovative application of transfer learning in crop type mapping, the methodology proposed in this study effectively addressed the problem of missing labels in target domains and alleviated the negative impact of domain shift. [ABSTRACT FROM AUTHOR]

Improved predictive models are needed in lung transplantation in the setting of a proposed allocation system that incorporates longer-term post-transplant survival in the United States. Allocation systems require accurate mortality predictions to justly allocate organs. Utilizing the United Network for Organ Sharing database (2005-2017), we fit models to predict 1-year mortality based on the Lung Allocation Score (LAS), the Chan, et al, 2019 model, a novel "clinician" model (a priori clinician selection of pre-transplant covariates), and two machine learning models (Least Absolute Shrinkage and Selection Operator; LASSO and Random Forests) for predicting 1-year and 3-year post-transplant mortality. We compared predictive accuracy among models. We evaluated the calibration of models by comparing average predicted probability vs observed outcome per decile. We repeated analyses fit for 3-year mortality, disease category, including donor covariates, and LAS era. The area under the cure for all models was low, ranging from 0.55 to 0.62. All exhibited reasonable negative predictive values (0.87-0.90), but the positive predictive value for was poor (all <0.25). Evaluating LAS calibration found 1-year post-transplant estimates consistently overestimated risk of mortality, with greater differences in higher deciles. LASSO, Random Forests, and clinician models showed no improvement when evaluated by disease category or with the addition of donor covariates and performed worse for 3-year outcomes. The LAS overestimated patients' risk of post-transplant death, thus underestimating transplant benefit in the sickest candidates. Novel models based on pre-transplant recipient covariates failed to improve prediction. There should be wariness in post-transplant survival predictions from available models. [ABSTRACT FROM AUTHOR]

Although several recent studies have examined psychosocial and demographic correlates of cannabis use disorder (CUD) in adults, few, if any, recent studies have evaluated the performance of machine learning methods relative to standard logistic regression for identifying correlates of CUD. The present study used pooled data from the 2015–2018 National Survey on Drug Use and Health to evaluate psychosocial and demographic correlates of CUD in adults. In addition, we compared the performance of logistic regression, classification trees, and random forest methods in classifying CUD. When comparing the performance of each method on the test data set, classification trees (AUC = 0.84, 95%CI: 0.82, 0.85) and random forest (AUC = 0.83, 95%CI: 0.82, 8.05) performed similarly and superior to logistic regression (AUC = 0.77, 95%CI: 0.74, 0.79). Results of the random forests reveal that marital status, risk propensity, age, and cocaine dependence variables contributed most to node purity, whereas model accuracy would decrease significantly if county type, income, race, and education variables were excluded from the model. One possible approach to improving the efficiency, interpretability, and clinical insights of CUD correlates is the employment of machine learning techniques. [ABSTRACT FROM AUTHOR]

• Dynamical S2S precipitation with a limited ensemble size leads to higher deterministic skill in predicting streamflow compared to ESP. • A ML technique was jointly applied with BCSD to improve S2S precipitation forecasts at watersheds in the United States. • The added skill of S2S precipitation propagates into streamflow predictions predominantly at longer forecast lead times. Accurate and reliable precipitation predictions made by dynamical forecast models could provide crucial information for human socioeconomic activities by enabling hydrologic forecasts at the Subseasonal-to-Seasonal (S2S) timescale. To utilize available S2S precipitation predictions for hydrologic forecasts, post-processing techniques have been applied to adapt the raw S2S precipitation to local watersheds. However, conventional statistical-based post-processing techniques are more focused on correcting the forecast bias, but rather limited in improving the predictive skill of available S2S precipitation forecasts. In this study, we combine the Random Forest classifiers (RF) with the Bias Correction and Spatial Disaggregation (BCSD) to adapt the 10-member ensemble precipitation forecast from the NASA Goddard Earth Observing System model version 5 (GEOS5) at 4 watersheds located in the NCEI South climate region of the United States. The adapted S2S precipitation is further applied for streamflow forecast by forcing a classical lumped hydrologic model. The performance of S2S precipitation as well as the corresponding streamflow predictions are benchmarked with the randomly resampled precipitation and the corresponding Ensemble Streamflow Prediction (ESP) framework-generated streamflow predictions. Evaluation statistics of Kling-Gupta Efficiency (KGE), Continuous Ranked Probability Skill Score (CRPSS), Reliability, Resolution, and Sharpness are employed to evaluate the predictive skill of precipitation and streamflow both deterministically and probabilistically. Our results indicate that dynamical S2S precipitation after forecast adaptation leads to consistently higher deterministic skill over ESP at all forecast lead times and across study watersheds. However, at longer forecast lead times beyond 10–15 days, S2S precipitation with a limited ensemble size does not present higher probabilistic skill than ESP. Our results shows that the joint application of RF and BCSD improves the predictive skill of the raw S2S precipitation at study watersheds in contrast to BCSD. Further, the added predictive skill of S2S precipitation brought by RF propagates into streamflow predictions, predominantly at longer forecast lead times exceeding 10 days. Overall, our results highlight the potential success of future work to apply other data-driven approaches to adapt the raw precipitation to local watersheds for more accurate and reliable streamflow forecasts at the S2S timescale. [ABSTRACT FROM AUTHOR]

Background Context: In 2008, the Centers for Medicare and Medicaid Services (CMS) established a list of hospital-acquired conditions (HACs) with significant deleterious effects on both patients and providers. Adult spinal deformity (ASD) surgery is complex and highly invasive, and as such may result in significant morbidity including these HACs.Purpose: Identify predictors for developing the most common HACs among adult spinal deformity (ASD) patients undergoing corrective surgery.Study Design/setting: Retrospective analysis.Patient Sample: One thousand one hundred and seventy-one ASD patients.Outcome Measures: HACs, Health-Related Quality of Life scores(HRQLs), Reoperation, Integrated Health State (IHS) METHODS: ASD pts undergoing surgery (>18 years, scoliosis ≥20°, SVA ≥5 cm, PT ≥25° and/or TK >60°) with complete data at BL and up to 2 years post-op were included. Patients were stratified by presence of >1 HAC, defined as at least one superficial/deep SSI, UTI, DVT, or PE within a 30-day post-op window. Random forest analysis generated 5,000 Conditional Inference Trees to compute a variable importance table for top predictors of HACs. An area-under-the-curve (AUC) methodology compared normalized HRQL scores between groups to determine an IHS with 2-year follow-up.Results: Total of 1,171 pts (59.8 years, 76.2%F, 28.1kg/m2) underwent corrective ASD surgery, with 1,053 pts in the non-HAC group and 118 in the HAC group. Of these pts, 25.4% had UTI, 15.4% DVT, 19.2% superficial SSI, 20.8% deep SSI, and 19.2% PE. HAC pts were on average older (63.5 vs 59.3, p=.004) and more often frail (51.3 vs 39.7%, p=.021) than non-HAC pts. Postop LOS and reoperation were most associated with HAC groups: [1] LOS >7 days [2] reoperation. Patient-related predictors of HACs were [3] age >50 yerr, [4] frailty, and [13] BMI >31. Procedure-related predictors of HACs were [5] operative-time >405 minutes, [6] levels fused >9, EBL >1450 mL, and [11] decompression. BL radiographic predictors were [7] PT >20°, [9] PI-LL>6°, [10] TL Cobb angle >15°, [12] SVA C7-S1 >29 mm. No differences were observed between groups with regards to IHS ODI (0.73 vs 0.74, p=.863), SRS (1.3 vs1.3, p=.374), NRS Back (0.6 vs 0.6, p=.158). HAC had higher rates of reoperation than non-HAC (0.08 vs 0.01, p=.066), and any HAC within 30-days of index was a significant predictor of reoperation (OR: 2.448 [1.94-3.09], p<.001).Conclusions: In a population of ASD patients, HACs were associated with length of stay, reoperation, age, and frailty. Radiographic parameters such as pelvic tilt >20°, PI-LL >6°, & SVA >29 mm also increased odds of HACs, and should raise postoperative awareness for HAC development. [ABSTRACT FROM AUTHOR]

Background: Transcatheter mitral valve repair (TMVR) utilization has increased significantly in the United States over the last years. Yet, a risk-prediction tool for adverse events has not been developed. We aimed to generate a machine-learning-based algorithm to predict in-hospital mortality after TMVR.Methods: Patients who underwent TMVR from 2012 through 2015 were identified using the National Inpatient Sample database. The study population was randomly divided into a training set (n = 636) and a testing set (n = 213). Prediction models for in-hospital mortality were obtained using five supervised machine-learning classifiers.Results: A total of 849 TMVRs were analyzed in our study. The overall in-hospital mortality was 3.1%. A naïve Bayes (NB) model had the best discrimination for fifteen variables, with an area under the receiver-operating curve (AUC) of 0.83 (95% CI, 0.80-0.87), compared to 0.77 for logistic regression (95% CI, 0.58-0.95), 0.73 for an artificial neural network (95% CI, 0.55-0.91), and 0.67 for both a random forest and a support-vector machine (95% CI, 0.47-0.87). History of coronary artery disease, of chronic kidney disease, and smoking were the three most significant predictors of in-hospital mortality.Conclusions: We developed a robust machine-learning-derived model to predict in-hospital mortality in patients undergoing TMVR. This model is promising for decision-making and deserves further clinical validation. [ABSTRACT FROM AUTHOR]

Despite a scarcity of potential donors for heart transplantation (HT) in the United States (US), a minority are actually accepted for HT. We evaluated donor characteristics associated with heart acceptance in the US and applied modern analytic methods to improve the prediction of heart acceptance. We included potential heart donors in the US from 2005 - 2020 (n = 73,948), a recent subset (n = 4,110, spanning 2015 - 2020) of which was enrolled in the Donor Heart Study (DHS). We identified and compared predictors of acceptance among DHS and other donors using logistic regression, incorporating interaction terms in the non-DHS ("nationwide") cohort to characterize time-varying effects. A prediction model was developed using prospectively-collected donor data in the DHS subset, and implemented in the form of a web-based prediction tool. Predictors of acceptance for HT were similar in the DHS and nationwide cohorts. A random forest model outperformed other prediction algorithms and the inclusion of previously unmeasured predictors (as captured in the DHS) improved model performance (AUC 0.90). In the nationwide cohort, older donor age has become more predictive of non-acceptance over the last 15 years while other factors - including mild cardiac imaging abnormalities, cocaine use, high troponin, hypertension, and Hepatitis C - have become less influential. Real-time prediction of donor heart acceptance may improve efficiency during donor management and allocation. The predictors of donor acceptance for HT have changed significantly over time, highlighting the need to continually re-evaluate and update our model. [ABSTRACT FROM AUTHOR]

• Roadside trees along the power lines are responsible for 90% of storm-related outages in the northeastern United States. • Spatially explicit vegetation risk models are necessary to guide the optimal targeting of vegetation management programs. • Remote sensing, LiDAR data in particular, plays a key role in modeling trees that are striking distance to powerlines. • LiDAR metrics, primary overhead length, and wire attributes emerged as the most influential variables when modeling tree risk at the device exposure zone level. Accurate vegetation risk modeling requires detailed and timely information on the physical structure of roadside trees in conjunction with local environmental and management factors. This study aims to understand the contribution of multiple environmental and management variables on tree-related power outages during storm events at a fine scale. We developed and compared five candidate vegetation risk models (VRMs) comprising 19 predictor variables falling into four categories: (1) forest characteristics, (2) soil and terrain, (3) vegetation management, and (4) utility infrastructure using Random Forest (RF) algorithm. The VRM model consisting of all four categories of predictor variables demonstrated the highest area under the receiver operating characteristic curve (AUC-ROC) value of 0.8320. The VRM that excluded tree-related variables exhibited the lowest AUC-ROC value of 0.7760. Two VRMs excluding soil/terrain and vegetation management variables performed marginally worse than the model with all variables. According to Shapley Additive Explanation analysis, the LiDAR-derived tree variables (proximity tree pixels), length of primary distribution overhead, median tree height, and wire properties (i.e., covered or bare) reported the highest importance-variables. This study's findings can guide vegetation risk at fine scales, which will help improve electrical grid resilience and hardening. [ABSTRACT FROM AUTHOR]

• Random forest models reveal seasonal signals in controls on river temperatures. • River temperatures are closely linked to climate but influenced by basin features. • Sensitivity of water temperatures to climate is strongly impacted by dams. Water temperature is vitally important to the health of rivers and streams, influencing the integrity of ecosystems, aquatic biogeochemistry, and the habitability of waterways for a variety of species. While climate is often regarded as the primary driver of stream temperature regimes, other factors - including hydrology, watershed characteristics, and human impacts - add substantial complexity to the variability of water temperatures. However, it remains challenging to disentangle the influence of these drivers through time and across rivers spanning diverse settings. To quantify the underlying controls on river thermal regimes, we applied conditional inference random forest models to predict maximum monthly stream temperatures and thermal sensitivities, averaged across a 4-year period (2016 to 2020), at 410 watersheds spanning the conterminous United States. Maximum stream temperatures were selected given their ecological relevance, while thermal sensitivity, which measures the relationship between air and water temperatures, was used to assess the responsiveness of stream temperatures to climate forcings. We interpreted these random forest models using variable importance rankings, describing seasonal and spatial variability in the dominant controls on water temperatures. Although our empirical results confirm that climate is indeed a primary control on temperature magnitude, our models highlight the diversity in drivers of water temperature variability across seasons, hydrologic regions, and between metrics. By combining random forest models with process-based understanding of stream thermal regimes, we provide new insights on the dynamic controls of water temperature variability across broad geographical domains, informing region- and season-specific controls for tailored thermal watershed management and guiding the framing of future water temperature modeling. [ABSTRACT FROM AUTHOR]

Abstract Particulate matter with aerodynamic diameter less than 2.5 μm (PM 2.5) is a complex mixture of chemical constituents emitted from various emission sources or through secondary reactions/processes; however, PM 2.5 is regulated mostly based on its total mass concentration. Studies to identify the impacts on climate change, visibility degradation and public health of different PM 2.5 constituents are hindered by limited ground measurements of PM 2.5 constituents. In this study, national models were developed based on random forest algorithm, one of machine learning methods that is of high predictive capacity and able to provide interpretable results, to predict concentrations of PM 2.5 sulfate, nitrate, organic carbon (OC) and elemental carbon (EC) across the conterminous United States from 2005 to 2015 at the daily level. The random forest models achieved high out-of-bag (OOB) R2 values at the daily level, and the mean OOB R2 values were 0.86, 0.82, 0.71 and 0.75 for sulfate, nitrate, OC and EC, respectively, over 2005–2015. The long-term temporal trends of PM 2.5 sulfate, nitrate, OC and EC predictions agreed well with their corresponding ground measurements. The annual mean of predicted PM 2.5 sulfate and EC levels across the conterminous United States decreased substantially from 2005 to 2015; while concentrations of predicted PM 2.5 nitrate and OC decreased and fluctuated during the study period. The annual prediction maps captured the characterized spatial patterns of the PM 2.5 constituents. The distributions of annual mean concentrations of sulfate and nitrate were generally regional in the extent that sulfate decreased from east to west smoothly with enhancement in California and nitrate had higher concentration in Midwest, Metro New York area, and California. OC and EC had regional high concentrations in the Southeast and Northwest as well as localized high levels around urban centers. The spatial patterns of PM 2.5 constituents were consistent with the distributions of their emission sources and secondary processes and transportation. Hence, the national models developed in this study could provide supplementary evaluations of spatio-temporal distributions of PM 2.5 constituents with full time-space coverages in the conterminous United States, which could be beneficial to assess the impacts of PM 2.5 constituents on radiation budgets and visibility degradation, and support exposure assessment for regional to national health studies at county or city levels to understand the acute and chronic toxicity and health impacts of PM 2.5 constituents, and consequently provide scientific evidence for making targeted and effective regulations of PM 2.5 pollution. Highlights • First national machine learning models to predict PM 2.5 constituents in the U.S. • Random forest models provide accurate and interpretable results with full coverage. • Random forest models explained over 70% of daily variability of PM 2.5 constituents. • Predictions capture both regional and localized patterns of PM 2.5 constituents. [ABSTRACT FROM AUTHOR]

Abstract Identification and prediction of facies and fractures are critical to subsurface geosystems analysis and hydrocarbon exploration. However, accurate prediction of facies and fractures is hard in the absence of high-resolution advanced geophysical logs (e.g. image logs) and core samples. This study demonstrates the application of machine learning algorithms, such as Bayesian Network Theory and Random Forest to predict the presence of different facies and fractures in sedimentary rocks using common well logs. We build and train supervised machine learning models using Bayesian Network Theory and Random Forest to classify facies and fractures in unconventional shale and conventional sandstone, and carbonate reservoirs. The trained machine-learning models are cross-validated 10-fold to check their robustness. Application of Bayesian Network theory and Random Forest shows that both facies and fractures can be predicted with high accuracy using limited common well logs. In addition, the Bayesian Network shows the complex causal relationship among the input petrophysical parameters and output (facies or fracture), which is its unique feature compared to other machine learning algorithms, such as Random Forest. Highlights • Application of machine learning algorithms for facies and fracture classification in conventional and unconventional reservoirs. • Bayesian Network provides an understanding of causality of relationships among petrophysical logs. • Random Forest provides a ranking of predictor variables. [ABSTRACT FROM AUTHOR]

• A combined physical and machine learning approach outperforms conventional approach to predict baseflow recession characteristics at ungauged stream locations. • Regions within the United States with similar baseflow recession characteristics are identified. • The proposed approach requires only basin area, stream length, and region identification to produce an estimate of baseflow recession characteristics at the ungauged stream location. Prediction of short- (i.e., aquifer is near or at saturated conditions) and long-time (i.e., aquifer is not near or at saturated conditions) baseflow recession characteristics at ungauged stream locations is a current challenge that has been primarily addressed by empirical approaches that relate these characteristics to basin attributes. However, the performance of these models is often only fair with coefficient of determination values ranging from 0.5 to 0.7. In this study, we propose a hybrid physical and machine learning approach to predict the long- and short-time baseflow recession characteristics at ungauged stream locations. This approach is compared to a machine learning method, random forest regression, that relates baseflow recession characteristics to basin attributes in 582 basins across the western and eastern United States. The new approach resulted in lower median and inner quartile ranges (IQR) of absolute normalized errors in predicting long-time baseflow recession characteristics (western: 23%, IQR=32%; eastern: 30%, IQR=39%) compared to estimates of those properties based on random forest regressions (western: 27%, IQR=34%; eastern: 38%, IQR=50%). For the short-time baseflow recession characteristics, the hybrid approach resulted in substantially lower median errors and IQR values (western: 79%, IQR=143%; eastern: 83%, IQR=140%) compared to estimates from random forest regressions (western: 1,577%, IQR=8,887%; eastern: 341%, IQR=2,154%). In addition, this approach identified four major regions in the western United States and three in the eastern United States where the baseflow recession characteristics are mostly constant, and these characteristics only vary based on the geometric properties of aquifers. Lastly, the inter-basin variability of the baseflow recession characteristics was not found to be strongly related to metrics measuring interstorm arrival periods, average number of storms, and average length of storms. [ABSTRACT FROM AUTHOR]

• Shape-based time-series clustering of gauged watersheds solely based on stream flow data. • Random Forests classification of ungauged watersheds to flow-regime classes. • Boruta algorithm identifies crucial watershed attributes used in Random Forests. • Flow-regime similarity does not equal to similarity in individual attributes. • Catchment Attributes and Meteorology for Large-Sample Studies Dataset. Watershed classification is considered necessary for various purposes, including improving the transferability of streamflow information and allowing the generalization of hydrologic theories. In this paper, we proposed a classification method using shape-based time-series clustering techniques to define hydrologically homogeneous classes for a set of 638 gauged watersheds in the conterminous United States. We defined 15 distinct flow-regime classes based on standardized weekly-step mean annual hydrographs of the watersheds analyzed. Most classes showed regionality, though to various degrees. Classes in the Atlantic Coast region showed strong geographical contiguity indicating that spatial proximity may be an appropriate indicator of flow-regime similarity; but on the other hand, nearby watersheds may exhibit different intra-annual variabilities and watersheds far apart from one another may exhibit a similar flow regime. Such cases show that spatial proximity may not be used as a universal indicator of flow-regime similarity. To expand the flow-regime classification to ungauged watersheds, where no streamflow data are available, we proposed a physical-characteristic classification method using the Random Forests algorithm based on a set of physical-climatic attributes. Based on the Boruta feature-selection algorithm, attributes related to snowiness, precipitation seasonality, and green vegetation coverage stood out as the most relevant controls on regime-class memberships. By analyzing the within-class variability of the watershed attributes, we found that watersheds of the same flow-regime class may be very different in terms of individual attribute values indicating that flow-regime dis/similarity may not be equated to dis/similarity in individual watershed attributes; instead, it is the interplay among attributes that determines the streamflow behavior. Overall, methods proposed for the classification of gauged and ungauged watersheds are believed to be of value in a range of applications including selecting gauged donor watersheds for estimating historical streamflow records at ungauged watersheds. [ABSTRACT FROM AUTHOR]

The global pandemic of SARS-CoV-2 (COVID-19) has been linked to adversely impacting individuals with opioid use disorder in the United States. This study focuses on analyzing opioid-involved mortality in the context of COVID-19 in the U.S. from a geospatial perspective. We investigated spatiotemporal patterns of opioid-involved deaths during 2020 and compared the spatiotemporal pattern of these deaths with patterns for the previous three years (2017–2019) to understand changes in the context of the COVID-19 pandemic. A counterfactual analysis framework together with a space-time random forest (STRF) model were used to estimate the increase in opioid-involved deaths related to the pandemic. To gain further insight into the relationship between opioid deaths and COVID-19-related factors, we built a space-time random forest model for the City of Chicago, that experienced a steep increase in opioid-related deaths during 2020. High ranking indicators identified by the model such as the number of positive COVID-19 cases adjusted by population and the change in stay-at-home dwell time during the pandemic were used to generate a vulnerability index for opioid overdoses during the COVID-19 pandemic in Chicago. • In 2020, opioid-involved deaths increased by 42.7% in the United States. • Analyzed the relationship between the COVID-19 pandemic and the number of opioid deaths in the U.S. • Applied a space-time random forest model in a counterfactual analysis framework. • Generated a ZIP-code level vulnerability index of opioid overdose occurrence during the pandemic for a metropolitan area. [ABSTRACT FROM AUTHOR]

• Cross-scale hydroclimatic properties are extensively investigated. • Temperature, precipitation and streamflow features are studied at nine time scales. • The spatial variability of the hydroclimate of the United States is assessed. • Hydroclimatic time series clustering is performed at nine temporal resolutions. • Hydroclimatic features are ranked according to their usefulness in clustering. A comprehensive understanding of the behaviours of the various geophysical processes and an effective evaluation of time series (else referred to as "stochastic") simulation models require, among others, detailed investigations across temporal scales. In this work, we propose a novel and detailed methodological framework for advancing and enriching such investigations in a hydroclimatic context. This specific framework is primarily based on a new feature compilation for multi-scale hydroclimatic analyses, and can facilitate largely interpretable feature investigations and comparisons in terms of temporal dependence, temporal variation, "forecastability", lumpiness, stability, nonlinearity (and linearity), trends, spikiness, curvature and seasonality. Multifaceted characterizations are herein obtained by computing the values of the proposed feature compilation across nine temporal resolutions (i.e., the 1-day, 2-day, 3-day, 7-day, 0.5-month, 1-month, 2-month, 3-month and 6-month ones) and three hydroclimatic time series types (i.e., temperature, precipitation and streamflow) for 34-year-long time series records originating from 511 geographical locations across the contiguous United States. Based on the acquired information and knowledge, similarities and differences between the examined time series types with respect to the evolution patterns characterizing their feature values with increasing (or decreasing) temporal resolution are identified. Moreover, the computed features are used as inputs to unsupervised random forests for detecting any meaningful clusters between the examined hydroclimatic time series. This clustering plays an illustrative role within this research, as it facilitates the identification of spatial patterns (with them consisting an important scientific target in hydroclimatic research) and their cross-scale comparison. We find that these specific patterns are largely analogous across temporal resolutions for the examined continental-scale region. We also apply explainable machine learning to compare the features with respect to their usefulness in clustering the time series at the various temporal resolutions. These latter investigations play a vital role within the proposed methodological framework, as they allow interpretation of hydroclimatic similarity at the various temporal resolutions. For most of the features, this usefulness can vary to a notable degree across temporal resolutions and time series types, thereby implying the need for conducting multifaceted time series characterizations for the study of hydroclimatic similarity. [ABSTRACT FROM AUTHOR]

• FWA in government administered programs is a major societal threat. • Analytical framework to convert claims data to meaningful fraud indicators. • Feature engineering can break or make a machine learning model. • Using engineered features keeps fraud in check and save billions of tax dollars. Insurance fraud is ranked second in the list of expensive crimes in the United States, with healthcare fraud being the second highest amongst all insurance fraud. Contrary to the popular belief, insurance fraud is not a victimless crime. The cost of crime is passed onto law-abiding citizens in the form of increased premiums or serious harm or danger to beneficiaries. To combat this kind of societal threat, there is an intense need for healthcare fraud detection systems to evolve. Some common roadblocks in implementing digital advancements (as seen in other domains) to healthcare are the complexity, heterogeneity of the data systems, and varied health program models across the United States. In other words, data are not stored in a centralized manner due to the sensitive domain nature, thus making it difficult to implement a robust real-world fraud-detection system. At the same time, in addition to the complexity of the varied systems involved, there is also the need to meet certain standards before a fraud actor can be prosecuted in a litigation setting. Thus, there is a human aspect to the fraud detection process flow in the real-world. In this article, a novel framework was outlined that converts diverse prescription claims (both fee-for-service and managed care) into a set of input variables/features suitable for implementation of an advanced statistical modeling fraud framework. This article thus aims to contribute to the existing literature by describing a process to transform prescription claims data to secondary features specific to provider fraud detection. The core idea was to focus on three main aspects of fraud (business heuristics on claims, provider-to-prescriber relations, and provider's client populations) to design the input features. A systematic method was proposed to extract features that have the potential to detect billing or behavioral outliers among pharmacy providers using information extracted from a secondary database (outpatient prescriptions). The application of a commonly used dimensionality reduction method, the Principal Component Analysis (PCA), was evaluated. PCA evaluates and reduces the extensive feature subspace to only those that captures the most variance in the data. To evaluate the features extracted from this framework, the application of the engineered features and the principal components to out-of-the-box logistic regression and Random Forest algorithms were considered to identify potential fraud. The engineered features when tested in different experimental settings with a logistic regression model had the highest area under the Receiver Operating Characteristic (ROC) curve of 0.76 and a weighted F score of 0.85 while a random forest model had the highest area under curve of 0.74 and a weighted F score of 0.88. [ABSTRACT FROM AUTHOR]

• More accurate and explanatory model is proposed for natural gas storage deliverability. • A hybrid Sparse Autoencoder and Random Forest model shows on average of 5.7% increase in prediction accuracy. • The first use of machine learning and explanatory methods (XAI) in deliverability prediction for natural gas storage. Numerous subsurface factors, including geology and fluid properties, can affect the connectivity of the storage spaces in depleted reservoirs; hence, fluid flow simulations become more complicated, and predicting their deliverability remains challenging. This paper applies Machine Learning (ML) techniques to predict the deliverability of underground natural gas storage (UNGS) in depleted reservoirs. First, three baseline models were developed based on Support Vector Regression (SVR), Artificial Neural Network (ANN), and Random Forest (RF) algorithms. To improve the accuracy of the RF model as the best-performing baseline model, a unified framework, referred to as SARF, was developed. SARF combines the capabilities of Sparse Autoencoder (SA) and that of Random Forest (RF). To achieve this, the internal representations of the SA, which constitute extracted features of the input variables, are used in RF to develop the proposed SARF framework. The predictive capabilities of the baseline models and the proposed SARF model were validated using 3744 real-world storage data samples of 52 active storage reservoirs in the United States. The experimental result of this study shows that the proposed SARF model achieved an average 5.7% increase in accuracy on four separate data partitions over the baseline RF model. Furthermore, a set of eXplainable Artificial Intelligence (XAI) methods were developed to provide an intuitive explanation of which factors influence the deliverability of reservoir storage. The visualizations developed using the XAI method provide an easy-to-understand interpretation of how the SARF model predicted the deliverability values for separate reservoirs. [ABSTRACT FROM AUTHOR]

Concerns of declining agrobiodiversity and widening socioeconomic inequities in United States (US) agriculture highlight the critical need for systemic change. Despite surmounting evidence of the field and landscape scale benefits of diversifying agricultural systems, path dependencies of US agriculture present barriers to such diversification pathways. This study aims to elucidate path dependencies of agricultural landscapes that (dis)incentivize crop diversification at the regional scale through two main research questions: 1) what are the biophysical and socioecological factors most predictive of agricultural diversity across the US; and 2) how do these factors vary regionally? Using a novel panel dataset constructed from several open-source databases, we use random forest (RF) permutation variable importance measures to identify and compare the factors most predictive of county-level crop diversity across nine US regions. Our results show that climate, land use norms, and farm inputs are consistently the most important categories for predicting agricultural diversity across regions; however, variability exists in the relative regional importance of variables within these categories. Thus, factors most strongly predictive of agricultural diversity across US landscapes operate distinctly at a regional level, emphasizing the need to consider multiple scales of influence. These distinct regional relationships contribute to path dependencies that present resistance to enhancing agricultural diversity. By more appropriately addressing the regional factors of US agricultural landscapes the constrain agricultural diversification, with an eye towards future cropscapes, we can shift current path dependencies toward a more resilient and adaptive US agricultural future. • Climate, land use norms, and farm inputs most strongly predict regional diversity. • Regional scale factors influence US agricultural path dependencies. • Reliance on chemical and fertilizer use does not promote diversification. [ABSTRACT FROM AUTHOR]

We asked two research questions: (1) What are the relative effects of climate change and climate-driven vegetation shifts on different components of future fire regimes? (2) How does incorporating climate-driven vegetation change into future fire regime projections alter the results compared to projections based only on direct climate effects? We used the western United States (US) as study area to answer these questions. Future (2071–2100) fire regimes were projected using statistical models to predict spatial patterns of occurrence, size and spread for large fires (> 400 ha) and a simulation experiment was conducted to compare the direct climatic effects and the indirect effects of climate-driven vegetation change on fire regimes. Results showed that vegetation change amplified climate-driven increases in fire frequency and size and had a larger overall effect on future total burned area in the western US than direct climate effects. Vegetation shifts, which were highly sensitive to precipitation pattern changes, were also a strong determinant of the future spatial pattern of burn rates and had different effects on fire in currently forested and grass/shrub areas. Our results showed that climate-driven vegetation change can exert strong localized effects on fire occurrence and size, which in turn drive regional changes in fire regimes. The effects of vegetation change for projections of the geographic patterns of future fire regimes may be at least as important as the direct effects of climate change, emphasizing that accounting for changing vegetation patterns in models of future climate–fire relationships is necessary to provide accurate projections at continental to global scales. [ABSTRACT FROM AUTHOR]

• A machine learning-based ensemble model was developed to quantify soil moisture (SM). • The ensemble model was effective in modeling dynamic SM over the CONUS. • The model showed robust performance over dense vegetation and complex topography. • The ensemble model yielded better performance than existing SM data. • Triple collocation analysis revealed the spatially robust performance of the model. Three widely used primary soil moisture (SM) data sources, namely, in-situ measurements, satellite observations, and land surface models (LSM), possess different characteristics. This study combined three SM data sources using machine learning (ML): random forest, artificial neural networks, and support vector regression, and simple averaging ensemble approaches to produce improved daily SM data over the contiguous United States (CONUS). For each ML model, three schemes were tested using different independent variables, namely, satellite-derived, LSM-derived, and both. Triple collocation analysis (TCA) was adopted to address the scale mismatch problem between in-situ and coarse gridded SM data. The proposed approach was evaluated using the International Soil Moisture Network (ISMN), Soil Moisture Active Passive Core Validation Sites (SMAP CVS), and TCA. In the ISMN-based evaluation, the proposed ML-based ensemble generally produced better evaluation metrics and showed robust skills over topographically complex and densely vegetated regions where existing SM products showed poor skills. The SMAP CVS-based evaluation demonstrated that the ML ensemble approach yielded a better performance than the existing SM datasets, resulting in a correlation coefficient of 0.78, unbiased root mean squared difference of 0.035 m3/m3, and bias of 0.006 m3/m3. In addition, the TCA results additionally confirmed that the ML-based ensemble had better spatiotemporal quality than the other SM products. The data-driven approach proposed in this study has three major novelties: (1) the proposed ML-based method synergistically merges various data sources to improve SM; (2) the performance of the proposed ML-based SM was robust to topography and vegetation; and (3) the average ensemble of three ML models additionally improves performances. The SM time-series data generated by the proposed approach are expectedly suitable variables for environmental and climate applications over CONUS. The research findings suggest that ML algorithms can be effectively used for modeling dynamic soil moisture. [ABSTRACT FROM AUTHOR]

[Display omitted] • Nationwide estimates of source-specific PM 2.5 mass at local scale is lacking. • Satellite measurements, land-use terms, and many variables were included as predictors. • We employed random forest algorithms with rigorous 10-fold cross validation. • Model predicts annual exposures of five major PM 2.5 sources at the census tract level for the entire United States. • Our estimates will help to examine source-specific health effects of air pollution in the national representative cohorts. Although PM 2.5 mass varies in source and composition over time and space, most health effects assessment have made the inherent assumption that all PM 2.5 mass has the same health implications, irrespective of composition. Nationwide estimates of source-specific PM 2.5 mass and constituents at local-scale would allow for epidemiological studies and health effects assessments that consider the variability in PM 2.5 characteristics in their health impact assessments. In response, we developed US models of annual exposures at the census tract level for five major PM 2.5 sources (traffic, soil, coal, oil, and biomass combustion) and six trace elements (elemental carbon, sulfur, silicon, selenium, nickel, and non-soil potassium) for 2001 through 2014. We employed Absolute Factor Analysis (APCA) to derive the source-specific PM 2.5 impacts at monitoring stations. Random forest algorithms that incorporated predictors derived from satellite, chemical transport model, and census tract resolution land-use data on traffic, meteorology, and emissions, which were rigorously tested by 10-fold cross-validation (CV), were then employed to estimate elemental and source-specific PM 2.5 levels at non-monitoring site census-tracts over the study years. Model performances were moderate to good, with CV R2 ranging from 0.41 to 0.95. For PM 2.5 sources, the highest CV R2 was attained for traffic PM 2.5 (CV R2 = 0.73), followed by coal (CV R2 = 0.65), oil (CV R2 = 0.62), soil (CV R2 = 0.60), and biomass (CV R2 = 0.41). Among constituents, the CV was highest for sulfur (CV R2 = 0.95). Our analyses provided highly resolved spatial estimates of annual elemental and source-specific PM 2.5 concentrations at the census-tract level, for 2001 through 2014. This dataset offers exposure estimates in support of future nationwide long-term health effects studies of source-specific PM 2.5 mass and constituents, enabling epidemiological research that addresses the fact that not all particles are the same. [ABSTRACT FROM AUTHOR]

Ensemble weather forecasts are often found to be under-dispersed and biased. Post-processing using spatio-temporal information is, therefore, required if one wishes to improve the quality of the raw forecasts. It is on this account that the present article generates and post-processes ensemble solar forecasts using satellite-derived irradiance not only from the focal pixel but also from the neighboring pixels. The ensemble forecasting model of choice is a dropout neural network with Monte Carlo sampling, eliminating the need for training multiple models and ensuring parameter diversity in ensemble forecasting. Subsequently, ensemble forecasts are post-processed using both parametric and nonparametric post-processing techniques, such as nonhomogenous regression, generalized additive model, linear quantile regression, or quantile random forests. The proposed forecasting framework is demonstrated and verified using four years of half-hourly data, at seven locations in the United States. Continuous ranked probability skill scores as high as 66% have been obtained when comparing the proposed method to a conditional climatology reference. The content of this article may be useful to a wide range of stakeholders in the power system, including but not limited to: independent system operators, who aim at efficiently maintaining the system's reliability; utility- and distributed-scale PV plant owners, who wish to avoid penalties for power deviation between the scheduled and real-time delivery; and forecast retailers, who can benefit from selling solar forecasts of higher quality. • Ensemble solar forecasts are generated using a dropout neural network with Monte Carlo sampling. • Solar forecasts are post-processed using information from neighboring satellite pixels. • Parametric and nonparametric methods are tested for ensemble post-processing. • The benefits offered by spatio-temporal information are investigated in a post-processing context. [ABSTRACT FROM AUTHOR]

Land surface temperature (LST) is a crucial parameter for hydrology, climate monitoring, and ecological and environmental research. LST products from thermal infrared (TIR) satellite data have been widely used for that. However, TIR information cannot provide LST data under cloudy-sky conditions. All-sky LST can be estimated from microwave measurements, but their coarse spatial resolution, narrow swaths, and short temporal range make it impossible to generate a long-term, high-resolution, accurate global all-sky LST global. This study proposes a methodology for generating the all-sky LST product by combining multiple data from Moderate Resolution Imaging Spectroradiometer (MODIS), reanalysis, and ground in situ measurements using a random forest. Field measurements from the AmeriFlux and Surface Radiation Budget (SURFRAD) networks were used for model training and validation. Cloudy-sky and clear-sky LST models were developed separately. To further improve the accuracy of the cloudy-sky LST model, the conventional RF model was extended to incorporate temporal information. The models were validated using in situ LST measurements from 2010, 2011, and 2017 that were not used for the model training. For the cloudy-sky and clear-sky models, root-mean-square-error (RMSE) = 2.767 and 2.756 K, R2 = 0.943 and 0.963, and bias = −0.143 and − 0.138 K, respectively. The same validation samples were used to validate both the MODIS LST product under clear-sky conditions and all-sky Global Land Data Assimilation System (GLDAS) LST product at 0.25° spatial resolution, with RMSE = 3.033 and 4.157 K, bias = −0.362 and − 0.224 K, and R2 = 0.904 and 0.955, respectively. Additionally, the 10-fold cross-validation results using all the training datasets further indicate the model stability. The models were applied to generate the all-sky LST product from 2000 to 2015 over the conterminous United States (CONUS). Our product shows similar spatial patterns to the MODIS and GLDAS LST products, but it is more accurate. Both validation and product comparisons demonstrated the robustness of our proposed models in generating the all-sky LST product. • A new scheme using RF method is developed to estimate all-sky LST by utilizing multisource data. • An improved RF method combing temporal information is proposed for the cloudy-sky model. • Our product shows similar spatial patterns to the MODIS and GLDAS LST products, but it is more accurate. • An all-sky LST data (2000–2015) is generated over the conterminous United States. [ABSTRACT FROM AUTHOR]

Commercial buildings account for a significant amount of total energy produced in the US, and the Heating Ventilation and Cooling (HVAC) systems are one of the most significant components of their overall consumption. In this study, we proposed a new data-driven approach to evaluate HVAC cooling systems in commercial buildings and identify savings opportunities. The focus is an investigation of the impact of thermostat setpoint setback but using only whole building, electricity data taken at 15-min intervals for the analysis. We conducted a comparative study of setpoint setback characteristics on 432 commercial buildings with 5 building usage types across the United States. To accomplish this, both piecewise and Random Forest regression algorithms were employed using electricity and exterior temperature datasets to identify operational characteristics and the effective setpoints in the building to determine the corresponding savings opportunities. Both occupied and unoccupied time periods were studied across cooling degree days (CDD), when air conditioning is typically operational. The results show that in commercial buildings, on average, cooling systems account for 9.5% of total consumption. When a one degree setback during the cooling season is applied, an average of approximately 1.1% of annual consumption is achieved; retail and office buildings demonstrate the highest potential for savings. Additionally, we identified that the number of cooling degree days and base to peak ratio (BPR) are the most important variables for predicting the magnitude of the consumption of cooling systems. • A new method for HVAC cooling thermostat setpoint setback savings is introduced. • The important variables in determination of HVAC cooling systems is discussed. • A comparative population study shows high potential of office and retail buildings. • One degree setback in air conditioning results in on average 1.1% annual savings. [ABSTRACT FROM AUTHOR]

Planned satellite launches will provide open access and operational L-band radar data streams at space-time resolutions not previously available. To further prepare, the Uninhabited Aerial Vehicle Synthetic Aperture Radar (UAVSAR) platform was used to observe cropland sites across the southern United States to support the development of L-band (24 cm) prototype science products. Time series flights flew over four independent areas during the growing season of 2019 while crop ground measurements were collected. Major crops include corn, cotton, pasture, peanut, rice, and soybean. A suite of cropland classification experiments applied a set of machine learning (ML) algorithms (random forest, feedforward fully connected neural network, support vector machine), the recently developed Multi-temporal Binary Tree Classification (MBTC), and a phenology (Coefficient of Variation; CoV) approach to synergistically assess performance, scattering mechanisms, and limitations. Specific objectives of this research application included 1.) evaluation of L-band mapping performance across multiple independent agricultural production areas with field scale training data, and 2.) assessment of the CoV approach for the generation of prototype NISAR Level 2 science products. Collectively, SAR terms with sensitivity to volume scattering performed well and consistently across CoV mapping experiments achieving accuracy greater than 80% for cropland vs not cropland. Dynamic phenology classes, such as herbaceous wetlands, had some confusion with CoV agriculture requiring further regionalized training optimization. Volume scattering and cross-pol terms were most useful across the different ML techniques with overall accuracy and Kappa consistently over 90% and 0.85, respectively, for crop type by late growth stages for L-band observations. As expected, time series information was more valuable compared to any single ML technique, site, or crop schema. Ultimately, as more SAR platforms launch, the user community should leverage physical contributions of different wavelengths and polarizations along with growing open access time series for efficient and meaningful agricultural products. • CoV and machine learning methods are deployed to understand L-band crop Monitoring. • CoV achieves an overall accuracy greater than 80% for the NISAR L2 product. • Time series and the cross-pol term are the most useful for discrimination. [ABSTRACT FROM AUTHOR]

• Random Forest model is used to predict annual commodity and virtual water flows. • Random Forest predictions consistently outperform those from a gravity model. • Model performance improves as commodities/regions are aggregated into coarser groups. • Model predicts both network connections and flows without comprising performance. • The modeled data could be beneficial for managing local and regional water scarcity. Virtual water flows are used to determine the indirect water requirements of a region or product, making them an indispensable tool for water sustainability analysis and assessment. Commodity flows are a key data needed to compute virtual water but are typically available every 5 years in the United States (US). The lack of continuous, annual commodity flow data severely limits our ability to study and understand the drivers, evolution, and alterations of virtual water in the US. We build and evaluate a machine learning model using Random Forest (RF) to predict annual commodity and virtual water flow networks. The model is used to perform several modeling experiments and illustrate the prediction of annual virtual water flows in the US during 2013–2018. We show that the RF predictions consistently outperform those from a gravity model. The overall performance of the RF algorithm improves as commodities or regions are aggregated into coarser groups. Likewise, the inclusion of past commodity flows as an additional explanatory variable enhances the RF performance. The combination of RF classification and regression allows predicting both network connections and flows without comprising performance. Based on our RF predictions for 2013–2018, we find that temporal variations in virtual water flows can be large for some regions in the US, underscoring the need addressed by this study of reconstructing domestic virtual water changes over time. By capturing inter-regional water consumption interactions in space and time, such reconstructed data could be beneficial in the future for anticipating and managing local and regional water scarcity. [ABSTRACT FROM AUTHOR]

The southeastern United States is unique in terms of both the intensity and scale of forest management, which includes substantial thinning and other forms of harvesting. Because thinning is not a land use transition, and the disturbance signal is relatively subtle compared to a clear cut, there is a dearth of studies that attempt to detect thinning over large areas. Our goal was to detect pine thins as an indicator of active forest management using Landsat data. Areas which undergo thinning are indicative of active forest management in the region. Our approach uses a machine learning method which combines first-order harmonics and metrics from 3-year Fourier regression of Landsat time series stacks, layers from the Global Forest Change product, and other vetted national products into a random forests model to classify forest thins in the southeastern US. Forest Harvest Inspection Records for Virginia were used for training and validation. Models were successful separating thins from clear cuts and non-harvested pines (overall accuracy 86%, clear cut accuracy 90%, thin accuracy 83% for a simplified 10-predictor variable model). Examination of variable importance illustrates the physical meaning behind the models. The curve fit statistics (R2 or RMSE) of the NDVI, Pan, and SWIR1 harmonic curve fits, which are an indication of a departure from typical vegetation phenology caused by thinning or other disturbances, were consistently among the top predictors. The harmonic regression constant, sine and cosine from the Landsat 8 panchromatic band were also important. These describe the visible reflectance (500–680 nm) phenology over the time period at a high spatial resolution (15 m). The Loss Year from the Global Forest Change product, which is an indication of stand replacing disturbance, was also consistently among the most important variables in the classifiers. High performance computing, such as Google Earth Engine, and analysis-ready data are important for this approach. This work has importance for quantification of actively managed forests in a region of the world where production forestry is the dominant land disturbance signal and a significant economic engine. • Demonstrated ability to separate thins from clear cuts and non-harvested pines. • Harmonic regression predictor variables improve classification of thins. • Panchromatic band of Landsat 8 is important for identification of forest thins. [ABSTRACT FROM AUTHOR]

Growing demands for temporally specific information on land surface change are fueling a new generation of maps and statistics that can contribute to understanding geographic and temporal patterns of change across large regions, provide input into a wide range of environmental modeling studies, clarify the drivers of change, and provide more timely information for land managers. To meet these needs, the U.S. Geological Survey has implemented a capability to monitor land surface change called the Land Change Monitoring, Assessment, and Projection (LCMAP) initiative. This paper describes the methodological foundations and lessons learned during development and testing of the LCMAP approach. Testing and evaluation of a suite of 10 annual land cover and land surface change data sets over six diverse study areas across the United States revealed good agreement with other published maps (overall agreement ranged from 73% to 87%) as well as several challenges that needed to be addressed to meet the goals of robust, repeatable, and geographically consistent monitoring results from the Continuous Change Detection and Classification (CCDC) algorithm. First, the high spatial and temporal variability of observational frequency led to differences in the number of changes identified, so CCDC was modified such that change detection is dependent on observational frequency. Second, the CCDC classification methodology was modified to improve its ability to characterize gradual land surface changes. Third, modifications were made to the classification element of CCDC to improve the representativeness of training data, which necessitated replacing the random forest algorithm with a boosted decision tree. Following these modifications, assessment of prototype Version 1 LCMAP results showed improvements in overall agreement (ranging from 85% to 90%). • We developed a robust capability for operational monitoring of land surface change. • Landsat ARD and Continuous Change Detection and Classification are foundational. • Landsat's rich time series has substantial variability in observation frequency. • The algorithm was modified reducing variability in results between scene centers and overlap zones. • Classification was modified to improve training data representativeness and reduce artifacts. [ABSTRACT FROM AUTHOR]