Simple Average Model Research Articles

Development of maximum relevant prior feature ensemble (MRPFE) index to characterize future drought using global climate models

Drought is one of the foremost outcomes of global warming and global climate change. It is a serious threat to humans and other living beings. To reduce the adverse impact of drought, mitigation strategies as well as sound projections of extreme events are essential. This research aims to strengthen the robustness of anticipated twenty-first century drought by combining different Global Climate Models (GCMs). In this article, we develop a new drought index, named Maximum Relevant Prior Feature Ensemble index that is based on the newly proposed weighting scheme, called weighted ensemble (WE). In the application, this study considers 32 randomly scattered grid points within the Tibetan Plateau region and 18 GCMs of Coupled Model Intercomparison Project Phase 6 (CMIP6) of precipitation. In this study, the comparative inferences of the WE scheme are made with the traditional simple model averaging (SMA). To investigate the trend and long-term probability of various classes, this research employs Markov chain steady states probability, Mann–Kendall trend test, and Sen’s Slope estimator. The outcomes of this research are twofold. Firstly, the comparative inference shows that the proposed weighting scheme has greater efficiency than SMA to conflate GCMs. Secondly, the research indicates that the Tibetan Plateau is projected to experience “moderate drought (MD)” in the twenty-first century.

Development of Ridge Ensemble Standardized Drought Index (RESDI) for improving drought characterization and future assessment.

Global warming upsets the environmental balance and leads to more frequent and severe climatic events. These extreme events include floods, droughts, and heatwaves. These widespread extreme events disrupt various sectors of ecosystems directly. However, among all these events, drought is one of the most prolonged climatic events that significantly destroys the ecosystem. Therefore, accurate and efficient assessment of droughts is necessary to mitigate their detrimental impacts. In recent years, several drought indices based on global climate models (GCMs) of Coupled Model Intercomparison Project Phase 6 (CMIP6) have been proposed to quantify and monitor droughts. However, each index has its advantages and limitations. As each index ensembles different models by using different statistical approaches, it is well known that the margin of error is always a part of statistics. Therefore, this study proposed a new drought index to reduce the uncertainty involved in the assessment of droughts. The proposed index named the Ridge Ensemble Standardized Drought Index (RESDI) is based on the innovative ensemble approach termed ridge parameters and distance-based weighting (RDW) scheme. And the development of this RDW scheme is based on two types of methods i.e., ridge regression and divergence-based method. In this research, we ensemble 18 different GCMs of CMIP6 using the RDW scheme. A comparative analysis of the RDW scheme is performed against the simple model average (SMA) and Bayesian model averaging (BMA) schemes at 32 locations on the Tibetan plateau. The comparison revealed that RDW has less mean absolute error (MAE) and root-mean-square error (RMSE). Therefore, the developed RESDI based on RDW is used to project drought properties under three distinct shared socioeconomic pathway (SSP) scenarios: SSP1-2.6, SSP2-4.5, and SSP5-8.5, across seven different time scales (1, 3, 7, 9, 12, 24, and 48). The projected data is then standardized by using the K-components Gaussian mixture model (K-CGMM). In addition, the study employs steady-state probabilities (SSPs) to determine the long-term behavior of drought. The outcome of this research shows that "normal drought (ND)" has the highest probability of occurrence under all scenarios and time scales.

Hybrid physics-machine learning models for predicting rate of penetration in the Halahatang oil field, Tarim Basin

Rate of penetration (ROP) is a key factor in drilling optimization, cost reduction and drilling cycle shortening. Due to the systematicity, complexity and uncertainty of drilling operations, however, it has always been a problem to establish a highly accurate and interpretable ROP prediction model to guide and optimize drilling operations. To solve this problem in the Tarim Basin, this study proposes four categories of hybrid physics-machine learning (ML) methods for modeling. One of which is residual modeling, in which an ML model learns to predict errors or residuals, via a physical model; the second is integrated coupling, in which the output of the physical model is used as an input to the ML model; the third is simple average, in which predictions from both the physical model and the ML model are combined; and the last is bootstrap aggregating (bagging), which follows the idea of ensemble learning to combine different physical models’ advantages. A total of 5655 real data points from the Halahatang oil field were used to test the performance of the various models. The results showed that the residual modeling model, with an R2 of 0.9936, had the best performance, followed by the simple average model and bagging with R2 values of 0.9394 and 0.5998, respectively. From the view of prediction accuracy, and model interpretability, the hybrid physics-ML model with residual modeling is the optimal method for ROP prediction.

Multivariate time series convolutional neural networks for long-term agricultural drought prediction under global warming

Agricultural drought (AD) is disastrous to crop production and plant growth. The prediction of AD with sufficient lead time is helpful for developing agricultural water strategy, particularly under the context of global warming. However, the previous studies mainly focused on short lead times (1–6 months) and only used 3 or less variables to predict AD through copula models. In this study, a novel multivariate time series convolutional neural network (T-CNN) is developed to predict AD with long lead times based on multiple meteorological variables. To demonstrate its feasibility and novelty, T-CNN is used in the Aral Sea Basin (ASB) where agricultural production is dominant. Three global climate models (GCMs) and three shared socioeconomic pathways (SSPs) from CMIP6 are considered during 2026–2100. Results indicate that (1) precipitation, temperature, potential evapotranspiration, relative humidity and northward wind are significantly correlated with AD, and are selected as the predictors of AD; (2) compared with the conventional CNN and convolutional long short-term memory (ConvLSTM), T-CNN’s performance is better, taking only about 10% of the computation time of ConvLSTM; (3) T-CNN can effectively extract the spatiotemporal characteristics of meteorological predictors and reproduce AD, showing high correlation coefficients (R>0.9) for 92.5% of the grids across ASB; (4) the result of simple model averaging (SMA) is better than other GCMs, indicating that the spatial differences in AD would become more pronounced with increasing time and emission level; (5) compared with the historical period, under SSP585, the extreme drought would increase 0.20 months/year (2026–2050), 0.23 months/year (2051–2075) and 0.28 months/year (2076–2100). The results highlight the spatiotemporal variation of AD in 21st century with a high resolution (0.1°×0.1°), which can provide scientific support for agricultural water management and long-term drought prevention in ASB.

Sales Predictor Using Simple Linear Regression and Simple Moving Average Model

Abstract: In the rapidly evolving business landscape, the accurate prediction of sales trends holds paramount importance for informed decision-making. Our project deals with the application of two fundamental forecasting models: SLR and SMA Models to predict future sales trajectories. The Simple Linear Average model involves the computation of a basic mean sales value, while the Simple Moving Average model incorporates a rolling average over a predefined time span. Upon successful training of the model, the model can predict the sales of the particular product with accuracy above 90%

A seismologically consistent surface rupture length model for unbounded and width-limited events

A new surface-rupture-length (SRL) relationship as a function of magnitude [Formula: see text], fault thickness, and fault dip angle is presented in this article. The objective of this study is to model the change in scaling between unbounded and width-limited ruptures. This is achieved through the use of seismological-theory-based relationships for the average displacement scaling and the aid of dynamic fault rupture simulations to constrain the rupture width scaling. The empirical data set used in the development of this relationship is composed of [Formula: see text] events ranging from [Formula: see text] 5 to [Formula: see text] and SRL 1.1 to 432 km. The dynamic rupture simulations data set includes [Formula: see text] events ranging from [Formula: see text] 4.9 to 8.2 and [Formula: see text] 1 to 655 km. For the average displacement [Formula: see text] scaling, three simple models and two composite models were evaluated. The simple average displacement models are a square root of the rupture area [Formula: see text], a down-dip width [Formula: see text], and a rupture length [Formula: see text] proportional model. The two composite models have a [Formula: see text] scaling for unbounded ruptures and transition to [Formula: see text] and [Formula: see text] scaling for width-limited events, respectively. The empirical data favors a [Formula: see text] scaling for the entire regime (unbounded and width-limited ruptures) followed by a [Formula: see text] scaling for unbounded that changes to [Formula: see text] scaling for width-limited ruptures. The selected models exhibit better predictive performance compared to linear [Formula: see text] type models, especially in the large magnitude range, which is dominated by width-limited events. A comparison with published SRL models shows consistent scaling for different fault types and tectonic environments.

A reduced-form ensemble of short-term air quality forecasting with the Sparrow search algorithm and decomposition error correction.

The level of air pollution is reflected by the air quality index (AQI). People can use the AQI to organize their activities in a way that reduces or prevents exposure to air pollution altogether. Based on the AQI, governments, organizations, and businesses can also make plans to reduce air pollution. The multi-model ensemble has recently become a popular method for forecasting time series; however, it encounters the research problems of multi-parameter optimization and interaction analysis. To this end, a reduced-form ensemble of short-term air quality forecasting with the Sparrow search algorithm and decomposition error correction model is proposed in this paper. First, the data are decomposed using the CEEMDAN decomposition algorithm. Second, the Sparrow search algorithm is used in the model training process to obtain the optimal hyperparameters of the deep learning model and construct the optimal deep learning model. Next, the constructed models are used to predict the decomposed data, and the Lagrange multiplier method is used to determine the weights of each deep learning model. At last, the prediction results of each deep learning model are combined according to the weights to obtain the combined prediction results. Experiments show that (1) GRU, Bi-GRU, LSTM, and Bi-LSTM are used to predict the undecomposed data and the data decomposed by CEEMADN. The outcomes demonstrate that the CEEMDAN decomposition technique can enhance the accuracy of the forecast, specifically an 11.248% reduction in average RMSE and a 0.865% increase in average R2. (2) A multi-model combination method based on the Lagrange multiplier method is designed, which can obtain the weights of each deep learning model, and the weights can combine multiple models. The results of the multi-model combination are better than those of the single model. (3) The Lagrange multiplier method was compared with the simple average combination model and the MAE inverse combination model. The experimental results show that the results obtained using the Lagrange multiplier method are better than the other two.

Temporal Upscaling of MODIS 1-km Instantaneous Land Surface Temperature to Monthly Mean Value: Method Evaluation and Product Generation

The monthly mean land surface temperature (MMLST) reflects more stable intra- and inter-annual temperature variations, and therefore, it has a wider range of applications than instantaneous land surface temperature (LST). This study aimed to generate a high-resolution global MMLST product by temporally upscaling the Moderate Resolution Imaging Spectroradiometer (MODIS) 1-km instantaneous LST. First, six current methods were comprehensively evaluated using cross-validation technology. These six methods are the cross combinations of two temporal aggregation schemes: the average by observations (ABO) and by days (ABD), and three conversion models: the diurnal temperature cycle model (DTC), a simple average model for two instantaneous LSTs (TSA), and a weighted average model for multiple instantaneous LSTs (MWA). The analysis with measurements from 235 flux stations worldwide revealed that the choice of conversion model considerably affected the overall retrieval accuracy whereas the influence of the aggregation scheme was minor. From the conversion model standpoint, MWA performed best, followed by DTC, and finally TSA; this order remained the same even if DTC and TSA were improved with mean bias correction. Notably, the errors of ABD <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MWA</sub> decreased as the number of daily mean LST (NOD) increased, whereas the errors of ABO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MWA</sub> were not related to NOD. Accordingly, we deduced that the optimal strategy for estimating MMLST is using ABO <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MWA</sub> when NOD is <20 and ABD <sub xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">MWA</sub> when NOD is ≥20. Subsequently, we adopted this combination method to process MODIS instantaneous LSTs and produced a global 1-km MMLST dataset for the years 2003–2020. The validation showed a satisfactory accuracy with a root mean square error of 1.6 K. The intercomparison with MMLSTs from geostationary satellites (containing complete LST daily cycle) presented a good agreement (biases < 0.3 K and STDs < 2 K). Compared with AIRS3STM product which had the same temporal span, the newly generated product exhibited a high consistency in reflecting temporal variations of global temperature. Most importantly, it had a prominently better ability to retrieve spatial details of temperature variations due to its higher resolution. Our new method and product show promising prospects for applications in global change studies where accurate spatially resolved MMLST data are one of the fundamental geophysical variables required.

From raw data to a score: comparing quantitative methods that construct multi-level composite implementation strength scores of family planning programs in Malawi

BackgroundData that capture implementation strength can be combined in multiple ways across content and health system levels to create a summary measure that can help us to explore and compare program implementation across facility catchment areas. Summary indices can make it easier for national policymakers to understand and address variation in strength of program implementation across jurisdictions. In this paper, we describe the development of an index that we used to describe the district-level strength of implementation of Malawi’s national family planning program.MethodsTo develop the index, we used data collected during a 2017 national, health facility and community health worker Implementation Strength Assessment survey in Malawi to test different methods to combine indicators within and then across domains (4 methods—simple additive, weighted additive, principal components analysis, exploratory factor analysis) and combine scores across health facility and community health worker levels (2 methods—simple average and mixed effects model) to create a catchment area-level summary score for each health facility in Malawi. We explored how well each model captures variation and predicts couple-years protection and how feasible it is to conduct each type of analysis and the resulting interpretability.ResultsWe found little difference in how the four methods combined indicator data at the individual and combined levels of the health system. However, there were major differences when combining scores across health system levels to obtain a score at the health facility catchment area level. The scores resulting from the mixed effects model were able to better discriminate differences between catchment area scores compared to the simple average method. The scores using the mixed effects combination method also demonstrated more of a dose–response relationship with couple-years protection.ConclusionsThe summary measure that was calculated from the mixed effects combination method captured the variation of strength of implementation of Malawi’s national family planning program at the health facility catchment area level. However, the best method for creating an index should be based on the pros and cons listed, not least, analyst capacity and ease of interpretability of findings. Ultimately, the resulting summary measure can aid decision-makers in understanding the combined effect of multiple aspects of programs being implemented in their health system and comparing the strengths of programs across geographies.

Multi-model ensemble approaches for simulation of evapotranspiration of karst agroforestry ecosystems

Water shortages frequently occur in karst areas, and there is an urgent need to quantify water fluxes to provide information for sustainable management of water resources. Thus, a variety of models have been developed to simulate the water balance process, including actual evapotranspiration (ETc), which is a key variable for linking water and energy cycles. However, high heterogeneity of the land surface makes it hard to get precise complete set of parameters for single model, and a single model often has uncertainties in simulating evapotranspiration in karst regions. Therefore, this study integrated three well-known individual models (Penman-Monteith, PM; Priestley and Taylor, PT; and Shuttleworth-Wallace, SW) with two multi-model ensemble approaches (Bayesian model averaging, BMA; and simple model average, SA) to enhance ETc modeling in a subtropical humid karst catchment. Results show that: 1) The individual models show different strengths for different ecosystems, which could be attributed to differences in the underlying landscape surface characteristics; 2) individual models exhibited seasonal uncertainties. For example, simulated ETc (ETs) by the PM and PT model was lower than ETo (observed ETc) during November-March but higher than during April-October for forest-grass mixed and grass ecosystems; 3) Two multi-model ensemble approaches (R2 ≥ 0.85) performed better than any individual model (R2 ≤ 0.85) most likely because multi-model ensemble approaches reduce model uncertainties by weakening the bias of individual models.

Mortality forecasting using stacked regression ensembles

Many alternative approaches for selecting mortality models and forecasting mortality have been proposed. The usual practice is to base forecasts on a single mortality model selected using in-sample goodness-of-fit measures. However, cross-validation measures are increasingly being used in model selection, and model combination methods are becoming a common alternative to using a single mortality model. We propose and assess a stacked regression ensemble that optimally combines different mortality models to reduce out-of-sample mean squared errors and mitigate model selection risk. Stacked regression uses a meta-learner to approximate horizon-specific weights by minimizing a cross-validation criterion for each forecasting horizon. The horizon-specific weights determine a mortality model combination customized to each horizon. We use 44 populations from the Human Mortality Database to compare the stacked regression ensemble with alternative methods. We show that, using one-year-ahead to 15-year-ahead out-of-sample mean squared errors, the stacked regression ensemble improves mortality forecast accuracy by 13% - 49% for males and 19% - 90% for females over individual mortality models. The stacked regression ensembles also have better predictive accuracy than other model combination methods, including Simple Model Averaging, Bayesian Model Averaging, and Model Confidence Set. We provide an R package, CoMoMo, that combines forecasts for Generalized-Age-Period-Cohort models.

Evaluation of Surface Upward Longwave Radiation in the CMIP6 Models with Ground and Satellite Observations

Surface upward longwave radiation (SULR) is an indicator of thermal conditions over the Earth’s surface. In this study, we validated the simulated SULR from 51 Coupled Model Intercomparison Project (CMIP6) general circulation models (GCMs) through a comparison with ground measurements and satellite-retrieved SULR from the Clouds and the Earth’s Radiant Energy System, Energy Balanced and Filled (CERES EBAF). Moreover, we improved the SULR estimations by a fusion of multiple CMIP6 GCMs using multimodel ensemble (MME) methods. Large variations were found in the monthly mean SULR among the 51 CMIP6 GCMs; the bias and root mean squared error (RMSE) of the individual CMIP6 GCMs at 133 sites ranged from −3 to 24 W m−2 and 22 to 38 W m−2, respectively, which were higher than those found between the CERES EBAF and GCMs. The CMIP6 GCMs did not improve the overestimation of SULR compared to the CMIP5 GCMs. The Bayesian model averaging (BMA) method showed better performance in simulating SULR than the individual GCMs and simple model averaging (SMA) method, with a bias of 0 W m−2 and an RMSE of 19.29 W m−2 for the 133 sites. In terms of the global annual mean SULR, our best estimation for the CMIP6 GCMs using the BMA method was 392 W m−2 during 2000–2014. We found that the SULR varied between 386 and 393 W m−2 from 1850 to 2014, exhibiting an increasing tendency of 0.2 W m−2 per decade (p &lt; 0.05).

Evaluation of crop model prediction and uncertainty using Bayesian parameter estimation and Bayesian model averaging

A recent trend in crop modeling has been the use of multi-model ensembles (MMEs) for impact assessment, especially as it relates to climate change. Studies have shown that, compared to individual models, the mean or median of a MME is a better predictor that is more accurate in making predictions and capable of providing model uncertainty information. In previous studies that used MMEs, each individual model was assigned an equal weight by simply averaging the predictions over all the models. Here we adopted a different method of creating MMEs, namely using Bayesian model averaging (BMA), which assigns different weights to individual models according to their performance of reproducing historical data. The main purpose of this study is to illustrate how BMA can be applied to crop models, and to compare its performance with simple model averaging. In addition, we also illustrate a full chain of analysis for crop model ensembles, which includes parameter estimation, ensemble creation and evaluation, uncertainty quantification and evaluation. The approaches were implemented with simulations of rice phenology using a small, three-member model ensemble. The results showed that the ensemble model e-BMA had nearly the same prediction accuracy as the best single model and that it predicts quite a bit better than the simply averaging ensemble named e-equal here. Squared bias was found to be the largest contributor to overall prediction uncertainty, both for the individual models and the ensemble models, and thus should be the first priority for model improvement. The estimated prediction uncertainties were larger than the variance of the observations. In the future, BMA should be tested more widely.

Evaluation of Subseasonal-to-Seasonal (S2S) precipitation forecast from the North American Multi-Model ensemble phase II (NMME-2) over the contiguous U.S.

The second phase of the North America Multi-Model Ensemble (NMME-2) provides globally available Subseasonal-to-Seasonal (S2S) precipitation forecasts with a daily resolution. The S2S precipitation forecasts are getting increasing attention for their potentials in providing hydrometeorological forcing information for water resources planning at an extended range. However, the forecast skills of many existing S2S forecast products will significantly decrease when the lead time increases, hindering their applicability for watershed-scale hydrologic modeling. Therefore, forecast validation and large-scale evaluation are of great importance for water resources planning and hydrological applications. In this study, we comprehensively evaluate the S2S precipitation forecasts from the NMME-2 dataset over the contiguous United States (CONUS) and during the study period from 1982 to 2011. Three aspects of precipitation forecast capabilities are compared and analyzed: bias, skill scores, and the ability to predict extreme precipitation events. The Parameter-elevation Regressions on Independent Slopes Model (PRISM) is used as ground truth reference. Differs from other regional forecast validation study, we further examined and analyzed the dependences of NMME-2 precipitation forecast skills according to different seasonality, geographical locations, and lead times. Results show that the forecast biases are not sensitive to lead times but are seasonally dependent of all NMME-2 models. Overestimations are found in the Western U.S. in cooler seasons while underestimations are observed in the central regions of the U.S. in warmer seasons. The forecast skill of all individual NMME-2 models generally decreases as increases of lead times. The simple model averaging (SMA) of five NMME-2 models demonstrates a higher forecast skill than any individual NMME-2 models. Spatially, the highest forecast skill scores are observed at coastal areas in the Western U.S. with an one-week lead time. As compared to the historical resampled forecasts, NMME-2 also shows better performance in predicting extreme precipitation events above 99% percentiles and below 1% percentiles with higher probability of detections and lower false alarm ratios. The obtained results suggest the great potentials of NMME-2 precipitation forecasts in assisting ensemble hydrologic forecasts at the S2S scale over the CONUS.

Using Deep Learning for Individual-Level Predictions of Adherence with Growth Hormone Therapy.

The problem of consistent therapy adherence is a current challenge for health informatics, and its solution can increase the success rate of treatments. Here we show a methodology to predict, at individual-level, future therapy adherence for patients receiving daily injections of growth hormone (GH) therapy for GH deficiency. Our proposed model is able to generate predictions of future adherence using a recurrent neural network with adherence data recorded by easypodTM, a connected autoinjection device. The model was trained with a multi-year long dataset with 2500 patients, from January 2007 to June 2019. When testing, the model reached an average sensitivity of 0.70 and a specificity of 0.88 per patient when predicting non-adherence (<85%) periods. When evaluated with thousands of therapy segments extracted from a test set, our model reached an AUC-PR score of 0.79 and AUC-ROC of 0.90; both metrics were consistently better than traditional approaches, such as simple average model. Using this model, we can perform precise early identification of patients who are likely to become non-adherent patients. This opens a path for healthcare practitioners to personalize GH therapy at any stage of the patients' journey and improve shared decision making with patients and caregivers to achieve optimal outcomes.

Analysis of Small Hydropower Generation Potential: (1) Estimation of the Potential in Ungaged Basins

Small hydropower (SHP) plants are advantageous as they have a short construction period and can be easily maintained. They also have a higher energy density than other alternative energy sources as environmentally-friendly energy sources. In general, hydropower potential is estimated based on the discharge in the river basin, and the discharge can be obtained from the stage station in the gaged basin. However, if there is no station (i.e., ungaged basin) or no sufficient discharge data, the discharge should be estimated based on rainfall data. The flow duration characteristic model is the most widely used method for the estimation of mean annual discharge because of its simplicity and it consists of rainfall, basin area, and runoff coefficient. Due to the characteristics of hydroelectric power depending on the discharge, there is a limit to guaranteeing the accuracy of estimating the generated power with only one method of the flow duration characteristic model. Therefore, this study assumes the gaged basins of the three hydropower plants of Deoksong, Hanseok, and Socheon in Korea exist as ungaged basins and the river discharges were simulated using the Kajiyama formula, modified-TPM(Two-Parameter Monthly) model, and Tank model for a comparison with the flow duration characteristics model. Furthermore, to minimize the uncertainty of the simulated discharge, four blending techniques of simple average method, MMSE(Multi-Model Super Ensemble), SMA(Simple Model Average), and MSE(Mean Square Error) were applied. As for the results, the obtained discharges from the four models were compared with the observed discharge and we noted that the discharges by the Kajiyama formula and modified-TPM model were better fitted with the observations than the discharge by the flow duration characteristics model. However, the result by the Tank model was not well fitted with the observation. Additionally, when we investigated the four blending techniques, we concluded that the MSE technique was the most appropriate for the discharge simulation of the ungaged basin. This study proposed a methodology to estimate power generation potential more accurately by applying discharge simulation models that have not been previously applied to the estimation of SHP potential and blending techniques were also used to minimize the uncertainty of the simulated discharge. The methodology proposed in this study is expected to be applicable for the estimation of SHP potential in ungaged basins.

On the use of machine learning based ensemble approaches to improve evapotranspiration estimates from croplands across a wide environmental gradient

Accurately mapping of regional-scale evapotranspiration (ET) from the croplands using remote sensing is currently challenged by limited spatial information on crop and field management to properly characterize the biophysical constraints on ET. A multi-model ensemble can potentially address this challenge, however, conventional ensemble models using the simple average (MEAN) or Bayesian Model Average (BMA) assign a fixed weight to each model and may not fully utilize the strengths of individual models. To this end, we developed four ensemble ET Models (EEMs) that use different machine learning (ML) classifiers, namely K-nearest neighbors, random forest, support vector machine, and multi-layer perception neural network (MLP), to assign varying weights to assemble six physically-driven remote sensing-based ET models. These ML-based EEMs were compared against the six individual ET models and two conventional ensemble methods (MEAN and BMA) using latent heat fluxes (λE) observations from 47 cropland eddy covariance flux sites covering diverse environments across the globe. Results suggested that while MEAN and BMA can reduce some uncertainties in the individual models, ML-based EEMs can better integrate the capabilities of multiple biophysical constraints on ET used across the individual models. The four ML-based EEMs yielded daily λE for training, validation, and testing datasets with the coefficient of determination (R2) and root mean squared error (RMSE) within 0.75 – 0.83 and 18 – 21 W m−2, respectively, among which the MLP algorithm was found to be the most efficient with respect to accuracies and costs. These performance metrics were much better than those from the conventional ensemble models (R2 = 0.69 – 0.71, RMSE = 23 – 25 W m−2) and six individual ET models (R2 = 0.53 – 0.69, RMSE = 26 – 35 W m−2). Results suggested that ML-based EEMs perform much better than the conventional approaches and hence can be viable tools for mapping cropland ET across a wide environmental gradient.

Applying probability calibration to ensemble methods to predict 2-year mortality in patients with DLBCL

BackgroundUnder the influences of chemotherapy regimens, clinical staging, immunologic expressions and other factors, the survival rates of patients with diffuse large B-cell lymphoma (DLBCL) are different. The accurate prediction of mortality hazards is key to precision medicine, which can help clinicians make optimal therapeutic decisions to extend the survival times of individual patients with DLBCL. Thus, we have developed a predictive model to predict the mortality hazard of DLBCL patients within 2 years of treatment.MethodsWe evaluated 406 patients with DLBCL and collected 17 variables from each patient. The predictive variables were selected by the Cox model, the logistic model and the random forest algorithm. Five classifiers were chosen as the base models for ensemble learning: the naïve Bayes, logistic regression, random forest, support vector machine and feedforward neural network models. We first calibrated the biased outputs from the five base models by using probability calibration methods (including shape-restricted polynomial regression, Platt scaling and isotonic regression). Then, we aggregated the outputs from the various base models to predict the 2-year mortality of DLBCL patients by using three strategies (stacking, simple averaging and weighted averaging). Finally, we assessed model performance over 300 hold-out tests.ResultsGender, stage, IPI, KPS and rituximab were significant factors for predicting the deaths of DLBCL patients within 2 years of treatment. The stacking model that first calibrated the base model by shape-restricted polynomial regression performed best (AUC = 0.820, ECE = 8.983, MCE = 21.265) in all methods. In contrast, the performance of the stacking model without undergoing probability calibration is inferior (AUC = 0.806, ECE = 9.866, MCE = 24.850). In the simple averaging model and weighted averaging model, the prediction error of the ensemble model also decreased with probability calibration.ConclusionsAmong all the methods compared, the proposed model has the lowest prediction error when predicting the 2-year mortality of DLBCL patients. These promising results may indicate that our modeling strategy of applying probability calibration to ensemble learning is successful.

An approach for multiclass skin lesion classification based on ensemble learning

Skin cancer is recognized as the most common kind of cancer in the world. It could be deadly if not identified at the primary stage, which makes early detection very crucial. It is possible to identify it with the naked eye, but high inter-class similarity and intra-class variations make it too hard to detect. Due to the prevalence of this disease around the world, so far many automated systems have been developed based on deep learning to assist the physician in the early detection of skin lesions. In this study, we propose a weighted average ensemble learning-based model to classify seven types of skin lesions. We used five deep neural network models, namely, ResNeXt, SeResNeXt, ResNet, Xception, and DenseNet as the base of the ensemble. For the training and evaluation of our models, we used 18,730 dermoscopy images collected from the official Human Against Machine (HAM10000) and ISIC 2019 dataset together with class balancing, noise removal, and data augmentation technique. We found the best combinations of the base models in the ensemble using the grid search method and optimized the impact of each base model for the average recall score. The five models performed excellent during evaluation and 88%, 89%, 91%, 88%, and 84% macro-average recall score were achieved by ResNeXt, SeResNeXt, DenseNet, Xception, and ResNet respectively. The simple average ensemble model boosted the result to 93% and the weighted average ensemble obtained a 94% recall score. The grid search method showed that the impact of all models are almost equal in the final model. The average ensemble can improve the result by a significant amount. Our proposed system performed better than other existing systems and can support dermatologists for diagnosis.

Mortality Forecasting Using Stacked Regression Ensembles

There are many alternative approaches to selecting mortality models and forecasting mortality. The standard practice is to produce forecasts using a single model such as the Lee-Carter, the Cairns-Blake-Dowd, or the Age- Period-Cohort model, with model selection based on in-sample goodness of fit measures. However, increasingly cross-validation measures based on forecasts are used in mortality model selection, and model combination methods such as Bayesian Model Averaging and Model Confidence Set have been proposed as alternatives to using a single model. We present a stacked regression ensemble method that optimally combines different mortality models to reduce the mean squared errors of mortality rate forecasts and mitigate model selection risk. Stacked regression uses a supervised machine learning algorithm to approximate the horizon-specific weights by minimizing the cross-validation criterion for each forecasting horizon. The horizon-specific weights facilitate the development of a mortality model combination customized to each horizon. Unlike other model combination methods, stacked regression simultaneously solves model selection and estimates model combinations to improve model forecasts. We use 44 populations from the Human Mortality Database to compare the stacked regression mortality approach with other alternative methods. We show that using one-year-ahead to 15−year-ahead out-of-sample mean squared errors, the stacked regression ensemble approach improves mortality forecast accuracy by 13% - 49% and 19% - 90% for females over individual mortality models. The stacked regression ensemble has better predictive accuracy than other model combination methods, including Simple Model Averaging, Bayesian Model Averaging, and Model Confidence Set. We provide a user-friendly open-source R package, CoMoMo, that estimates models and provides mortality rate forecasts for different mortality model combination methods.