No full-text available
To read the full-text of this research,
you can request a copy directly from the authors.
Accurate real-time monitoring of fine dust using a densely connected convolutional networks with measured plasma emissions
Abstract
Accurate identification and monitoring of fine dust are emerging as a primary global issue for addressing the harmful effects of fine dust on public health. Identifying the source of fine dust is indispensable for ensuring the human lifespan as well as preventing environmental disasters. Here a simple yet effective spark-induced plasma spectroscopy (SIPS) unit combined with deep learning for real-time classification is verified as a fast and precise PM (particulate matter) source identification technique. SIPS promises portable use, label-free detection, source identification, and chemical susceptibility in a single step with acceptable speed and accuracy. In particular, the densely connected convolutional networks (DenseNet) are used with measured spark-induced plasma emission datasets to identify PM sources at above 98%. The identification performance was compared with other common classification methods, and DenseNet with dropouts (30%), optimized batch size (16), and cyclic learning rate training emerged as the most promising source identification method.
... Reference (Yang et al. 2022) utilized a densely connected convolutional network to measure the plasma emissions for accurate real-time dust monitoring. Although the densely connected convolutional network can process dust image data and provide dust prediction results, the results are usually presented in the form of numerical values or probabilities. ...
... The accuracy of monitoring results directly reflects the accuracy of intelligent monitoring of dust pollution. Then, the method proposed in this paper is compared with the methods in references (Hyung et al. 2024;Martha et al. 2023;Jamei et al. 2022), and (Yang et al. 2022) to test the problems of different methods in dust monitoring. Using accuracy and recognition time as experimental indicators, five methods were used to monitor data images on sunny and sandstorm days, and the comparison results are shown in Table 2. ...
In order to realize fast, remote, and accurate intelligent monitoring of dust pollution in the construction of bare land areas, an intelligent monitoring method for unmanned aerial vehicles (UAVs) for environmental protection of dust pollution in the construction of bare land areas based on lightweight improvement YOLOv5 is studied. The image of the construction bare land area is remotely collected by the camera carried by the UAV, and the image enhancement algorithm for intelligent monitoring of the construction bare land area based on the bistable system is used to map the image gray value to the bipolar aperiodic pulse amplitude modulation signal. The resonance characteristics of the bistable stochastic resonance system are used to suppress the noise and enhance the image. As a recognition sample for the intelligent recognition model of fugitive dust images in construction bare land areas, which is based on the lightweight-improved YOLOv5. This model uses the improved coordinated attention, and the backbone network of coordinated attention (CA) mechanism extracts the important features in the monitoring images of bare areas before and after the increase and uses the weighted bidirectional feature pyramid improved by the network neck to achieve multi-scale feature fusion. Subsequently, the fusion-extracted feature map is used to identify the boundary box of dust targets and realize the intelligent monitoring of dust environmental protection UAV in the construction of bare areas. Through testing, owing to the lightweight improvement of the YOLOv5 network, the intelligent monitoring results of unmanned aerial vehicles for fugitive dust and environmental protection in the construction bare land area are accurate and efficient.
Graphical Abstract
... The classical residual network architecture adds the inputs of a layer stack to its outputs and then passes the result to the next stack [6].The DenseNet architecture extends this principle by introducing additional connections from the output of each stack, to the inputs of every other subsequent stack. The above approaches have demonstrated significant advantages in improving the trainability of deeper architectures [7]. N-BEATS, on the other hand, uses a dual residual residual structure, as shown in the middle and right 2 parts of Fig. 1, which has two residual branches, one running on the backcast prediction at each layer, and the other on the forecast branch at each layer, to better utilize the advantages of residual networks. ...
The water level of hydrological observatory has important reference significance in water resources utilization, drought and flood control and water supply and drainage. Aiming at the problem of poor accuracy and difficult training in long-term water level prediction, an improved time series model, WOA-N-Beats, is proposed, and the whale optimization algorithm is introduced to solve the problem of the model falling into local optimal solutions and the difficulty of finding optimal parameters. In the water level prediction of Wulong, Yichang, and Shashi stations in the next 512 days, the experiments show that compared with RNN and LSTM methods, WOA-N-Beats has the best MAE, MSE, and RMSE, and the enhancement effect is large, which realizes a more accurate prediction.
The goal of music genre classification is to identify the genre of given feature vectors representing certain characteristics of music clips. In addition, to improve the accuracy of music genre classification, considerable research has been conducted on extracting spectral features, which contain critical information for genre classification, from music clips and feeding these features into training models. In particular, recent studies argue that classification accuracy can be enhanced by employing multiple spectral features simultaneously. Consequently, fusing information from multiple spectral features is a critical consideration in designing music genre classification models. Hence, this paper provides a short survey of recent studies on music genre classification and compares the performance of the most recent CNN-based models with a newly devised model that employs a late fusion strategy for the multiple spectral features. Our empirical study of 12 public datasets, including Ballroom, ISMIR04, and GTZAN, showed that the late fusion CNN model outperforms other compared methods. Additionally, we performed an in-depth analysis to validate the effectiveness of the late fusion strategy in music genre classification.
With the increasing problem of atmospheric particulate matter (PM) pollution, an increasing number of scholars has conducted studies related to PM. Most of the research on PM has focused on information measurement such as particle size, where light scattering method is widely used in the field of PM measurement. Numerous studies have shown that the particle properties have a great influence on the measurement accuracy of light scattering method. However, there are relatively few studies related to the identification of particle properties. Therefore, according to the difference of particle signals with different properties, this article proposes a new method of particle property identification based on probabilistic neural network (PNN). Also, a new portable light scattering sensor integrated with virtual impactor (VI) is designed for particle signal acquisition. In the particle signal feature extraction session, we innovatively select Image entropy, kurtosis, and sample entropy of discrete wavelet approximation signal as eigenvalues of the particle signal. The experimental results show that the VI has a good separation with a cutting particle size of 2.67
. The designed new portable light scattering sensor has good performance which can realize the task of particle light scattering signal acquisition. Meanwhile, the proposed new method of particle property identification has a 99% correct classification rate. In future work, we will further optimize the design of portable sensors and thus reduce the specifications of the measurement system.
This study explored the toxic gas detection using a set of two spark-induced plasma emission spectroscopy (SIPS) modules as an alternative to conventional chemical sensors. This opens up a new possibility for detecting toxic molecules (formaldehyde, acetaldehyde, acetic acid, toluene, and ammonia) in real-time (<3 s) at relatively high sensitivity (<5 ppm). An optimized electrical controller (Raspberry Pi), manufactured as a compact and economical 3-channel optical measuring device, was developed for handling the high-resolution time-resolved electrical signals from the plasma emissions. Subsequently, the findings of this research elucidate the usability of the 3-channel SIPS device for quantitative monitoring of toxic gases.
Recently, heavy metal air pollution has received significant interest in computing the total concentration of every toxic metal. Chemical fractionation of possibly toxic substances in airborne particles becomes a vital element. Among the primary and secondary air pollutants, airborne particulate matter (APM) received considerable internet among research communities owing to the adversative impact on human health. Hence, size distribution details of airborne heavy metals are important in assessing the adverse health effects over the globe. Recently, deep learning models have gained significant interest over the mathematical and statistical prediction models. In this view, this paper presents a novel arithmetic optimization algorithm (AOA) with multi-head attention based bidirectional long short-term memory (MABLSTM) model for predicting the size fractionated airborne particle bound metals. The proposed AOA-MABLSTM technique focuses on the forecasting of the size-fractionated airborne particle bound matter. The presented model intends to examine the concentration of PM and distinct sized-fractionated APM. The proposed model establishes MABLSTM based accurate predictive approaches for atmospheric heavy 83 metals is used for determining temporal trend of heavy metal. The proposed model employs AOA based hyperparameter tuning process to optimally tune the hyperparameters included in the MABLSTM method. To demonstrate the improved outcomes of the AOA-MABLSTM approach, a comparison study is performed with recent models. The stimulation results emphasized the betterment of the presented model over the other methods. Aluminum metal had an RMSE of 73.200 for AOA-MABLSTM. On Cu metal, the AOA-MABLSTM approach had an RMSE of 6.747. On Zn metal, the AOA-MABLSTM system lowered the RMSE by 45.250.
This study investigated the chemical profiles of fine urban road dust as a set of indicators for major air pollutants at sampling sites or as proxies for potential human health impacts. We examined the chemical compositions of fine particles (< 100 μm) or re-suspended ultrafine particles (< 2.5 μm) in the urban road dust collected from the cities with major emission sources of CO, NH3, NOx, PM2.5, SOx, and volatile organic compounds. The elemental compositions, including metal contents and volatile or semi-volatile organic compound species were determined to constitute comprehensive chemical profiles of the solid road dust samples. The water-extractable organic compounds and fluorescent species of the size-fractionated re-suspended fine particulate matter (RPM) were also incorporated in the chemical profiles. The metal content and aliphatic hydrocarbons could partly distinguish emission sources, and clearer distinctions were achieved with the inclusion of fluorescence excitation-emission matrix (EEM) results. The dose–response test results showed positive correlations between cytotoxicity and relative abundance of hydrocarbons or metal contents of urban road dust. The set of chemical profiles suggested in this study could be further utilized for site identification or human health impact assessment using urban road dust.
The dust dispersed composition in various types of hard and soft works is under study. The dust emissions indoors concentration data is obtained. The dust concentration on distance dependence is constructed. The impact assessment of hard and soft works in warehouses and residential premises on environmental pollution by dust emissions is given. The concentration and ratio of PM2.5 to PM10 particles were calculated in order to determine the dustiness and the negative impact on the environment.
Determination of trace elements in soils with laser-induced breakdown spectroscopy is significantly affected by the matrix effect, due to large variations in chemical composition and physical property of different soils. Spectroscopic data treatment with univariate models often leads to poor analytical performances. We have developed in this work a multivariate model using machine learning algorithms based on a back-propagation neural network (BPNN). Beyond the classical chemometry approach, machine learning, with tremendous progresses the last years especially for image processing, is offering an ensemble of powerful and constantly renewed algorithms and tools efficient for the different steps in the construction of a spectroscopic data treatment model, including feature selection and neural network training. Considering the matrix effect as the focus of this work, we have developed the concept of generalized spectrum, where the information about the soil matrix is explicitly included in the input vector of the model as an additional dimension. After a brief presentation of the experimental procedure and the results of regression with a univariate model, the development of the multivariate model will be described in detail together with its analytical performances, showing average relative errors of calibration (REC) and of prediction (REP) within the range of 5–6%.
In this study, 121 daily PM2.5 (aerosol particle with aerodynamic diameter less than 2.5 μm) samples were collected from an urban site in Beijing in four months between April 2009 and January 2010 representing the four seasons. The samples were determined for various compositions, including elements, ions, and organic/elemental carbon. Various approaches, such as chemical mass balance, positive matrix factorization (PMF), trajectory clustering, and potential source contribution function (PSCF), were employed for characterizing aerosol speciation, identifying likely sources, and apportioning contributions from each likely source. Our results have shown distinctive seasonality for various aerosol speciations associated with PM2.5 in Beijing. Soil dust waxes in the spring and wanes in the summer. Regarding the secondary aerosol components, inorganic and organic species may behave in different manners. The former preferentially forms in the hot and humid summer via photochemical reactions, although their precursor gases, such as SO2 and NOx, are emitted much more in winter. The latter seems to favorably form in the cold and dry winter. Synoptic meteorological and climate conditions can overwhelm the emission pattern in the formation of secondary aerosols. The PMF model identified six main sources: soil dust, coal combustion, biomass burning, traffic and waste incineration emission, industrial pollution, and secondary inorganic aerosol. Each of these sources has an annual mean contribution of 16, 14, 13, 3, 28, and 26%, respectively, to PM2.5. However, the relative contributions of these identified sources significantly vary with changing seasons. The results of trajectory clustering and the PSCF method demonstrated that regional sources could be crucial contributors to PM pollution in Beijing. In conclusion, we have unraveled some complex aspects of the pollution sources and formation processes of PM2.5 in Beijing. To our knowledge, this is the first systematic study that comprehensively explores the chemical characterizations and source apportionments of PM2.5 aerosol speciation in Beijing by applying multiple approaches based on a completely seasonal perspective.
In this study, a method based on laser-induced breakdown spectroscopy (LIBS) was developed to detect soil contaminated with Pb. Different levels of Pb were added to soil samples in which tobacco was planted over a period of two to four weeks. Principal component analysis and deep learning with a deep belief network (DBN) were implemented to classify the LIBS data. The robustness of the method was verified through a comparison with the results of a support vector machine and partial least squares discriminant analysis. A confusion matrix of the different algorithms shows that the DBN achieved satisfactory classification performance on all samples of contaminated soil. In terms of classification, the proposed method performed better on samples contaminated for four weeks than on those contaminated for two weeks. The results show that LIBS can be used with deep learning for the detection of heavy metals in soil.
Abstract Fine particulate matters less than 2.5 µm (PM2.5) in the ambient atmosphere are strongly associated with adverse health effects. However, it is unlikely that all fine particles are equally toxic in view of their different sizes and chemical components. Toxicity of fine particles produced from various combustion sources (diesel engine, gasoline engine, biomass burning (rice straw and pine stem burning), and coal combustion) and non-combustion sources (road dust including sea spray aerosols, ammonium sulfate, ammonium nitrate, and secondary organic aerosols (SOA)), which are known major sources of PM2.5, was determined. Multiple biological and chemical endpoints were integrated for various source-specific aerosols to derive toxicity scores for particles originating from different sources. The highest toxicity score was obtained for diesel engine exhaust particles, followed by gasoline engine exhaust particles, biomass burning particles, coal combustion particles, and road dust, suggesting that traffic plays the most critical role in enhancing the toxic effects of fine particles. The toxicity ranking of fine particles produced from various sources can be used to better understand the adverse health effects caused by different fine particle types in the ambient atmosphere, and to provide practical management of fine particles beyond what can be achieved only using PM mass which is the current regulation standard.
Source apportionment analysis of hourly resolved particulate matter (PM) speciation data was performed using positive matrix factorization (PMF). The data were measured at an urban site in downtown Toronto, Canada during two campaign periods (April-July, 2013; November, 2013-February, 2014), and included trace metals, black carbon, and mass 10 spectra for organic and inorganic species (PMFFull). The chemical composition was measured by collocated high time resolution instrumentation, including an Aerosol Chemical Speciation Monitor, an Xact metals monitor, and a seven-wavelength Aethalometer. Separate PMF analyses were conducted using the trace metal only data (PMFmetal) and organic mass spectra only (PMForg), and compared with the PMFFull results. Comparison of these three PMF analyses demonstrated that the full analysis offered many advantages in the apportionment of local and regional sources compared to using the organic or metals data 15 individually. In combining the high time resolution data, this analysis enabled i) the quantification of metal-rich sources of PM2.5 (PM < 2.5 μm), ii) the resolution of more robust factor profiles and contributions, and iii) the identification of additional organic aerosol sources. Nine factors were identified through the PMFFull analysis: five local factors (i.e. Road Dust, Primary Vehicle Emissions, Tire Wear, Cooking, and Industrial Sector) and four regional factors (i.e. Biomass Burning, Oxidised Organics, Sulphate and 20 Oxidised Organics, and Nitrate and Oxidised Organics). The majority of the metal emissions (83%) and almost half of the black carbon (49%) were associated with the three traffic-related factors which, on average, contributed a minority (17%) of the overall PM2.5 mass. Strong seasonal patterns were observed for the traffic-related emissions: higher contributions of resuspended road dust in spring vs. a winter high for tire wear related emissions. Biomass Burning contributed the majority of the PM2.5 mass (52%) in June and July due to a major forest fire event. Much of this mass was due to photochemical aging of the biomass 25 burning aerosol. On average, industrially related factors contributed almost half (49%) of the PM2.5; most of this mass was secondary aerosol species. Nitrate coupled with highly oxidised organics was the largest contributor, accounting for 30% of PM2.5 on average, with higher levels in winter and at night. Including the temporal variabilities of inorganic ions and trace metals in the PMFFull analysis provided additional structure to subdivide the low volatility oxidised organic aerosol into three sources. Resuspended road dust was identified as a potential source of aged organic aerosol. 30 The novelty of this study is the application of PMF receptor modeling to hourly resolved trace metals in conjunction with organic mass spectra, inorganic species, and black carbon for different seasons, and the comparison of separate PMF analyses applied to metals or organics alone. The inclusion of these different types of hourly data allowed more robust apportionment of PM sources, as compared to analysing organic or metals data individually.
Recent work has shown that convolutional networks can be substantially deeper, more accurate and efficient to train if they contain shorter connections between layers close to the input and those close to the output. In this paper we embrace this observation and introduce the Dense Convolutional Network (DenseNet), where each layer is directly connected to every other layer in a feed-forward fashion. Whereas traditional convolutional networks with L layers have L connections, one between each layer and its subsequent layer (treating the input as layer 0), our network has L(L+1)/2 direct connections. For each layer, the feature maps of all preceding layers are treated as separate inputs whereas its own feature maps are passed on as inputs to all subsequent layers. Our proposed connectivity pattern has several compelling advantages: it alleviates the vanishing gradient problem and strengthens feature propagation; despite the increase in connections, it encourages feature reuse and leads to a substantial reduction of parameters; its models tend to generalize surprisingly well. We evaluate our proposed architecture on five highly competitive object recognition benchmark tasks. The DenseNet obtains significant improvements over the state-of-the-art on all five of them (e.g., yielding 3.74% test error on CIFAR-10, 19.25% on CIFAR-100 and 1.59% on SVHN).
The aim of this paper is the description of the optical emission spectral features of the plasma
produced by laser–matter interaction, from a fundamental point of view. The laser induced
plasma emission spectra are discussed in connection with the basic mechanisms that take
place in the plasma phase at different time delays from the laser pulse. The laser induced
plasma being a dynamic system, the hierarchy of the elementary mechanisms changes
continuously because the electron number density and the electron temperature decrease
during the expansion. As a consequence of this, over the duration of the plasma’s persistence
the prevailing emitting species changes from ions to atoms and from atoms to molecules. Both
atomic and molecular emission spectroscopy are discussed, to convey a complete description
of the temporal evolution of laser induced plasma.
Current literature, as well as the traditional plasma theories, are presented and discussed
in order to give to the reader a general idea of the potentialities and drawbacks of emission
spectroscopy in the study of laser induced plasma and its various applications.
Many studies have demonstrated that heavy metals existing as a mixture in the atmospheric environment cause adverse effects on human health and are important key factors of cytotoxicity; however, little investigation has been conducted on a toxicological study of a metal mixture from atmospheric fine particulate matter. The objective of this study was to predict the combined effects of heavy metals in aerosol by using in vitro human cells and obtain a suitable mixture toxicity model. Arsenic, nickel, and lead were selected for mixtures exposed to A549 human lung cancer cells. Cell proliferation (WST-1), glutathione (GSH), and interleukin (IL)-8 inhibition were observed and applied to the prediction models of mixture toxicity, concentration addition (CA) and independent action (IA). The total mixture concentrations were set by an IC10-fixed ratio of individual toxicity to be more realistic for mortality and enzyme inhibition tests. The results showed that the IA model was statistically closer to the observed results than the CA model in mortality, indicating dissimilar modes of action. For the GSH inhibition, the results predicted by the IA and CA models were highly overestimated relative to mortality. Meanwhile, the IL-8 results were stable with no significant change in immune reaction related to inflammation. In conclusion, the IA model is a rapid prediction model in heavy metals mixtures; mortality, as a total outcome of cell response, is a good tool for demonstrating the combined toxicity rather than other biochemical responses.
During winter 2013–2014 aerosol mass spectrometer (AMS) measurements were
conducted for the first time with a novel PM2.5 (particulate matter
with aerodynamic diameter ≤ 2.5 µm) lens in two major cities of
China: Xi'an and Beijing. We denote the periods with visibility below 2 km
as extreme haze and refer to the rest as reference periods. During the
measurements in Xi'an an extreme haze covered the city for about a week and
the total non-refractory (NR)-PM2.5 mass fraction reached peak
concentrations of over 1000 µg m−3. During the measurements in
Beijing two extreme haze events occurred, but the temporal extent and the
total concentrations reached during these events were lower than in Xi'an.
Average PM2.5 concentrations of 537 ± 146 and
243 ± 47 µg m−3 (including NR species and equivalent black
carbon, eBC) were recorded during the extreme haze events in Xi'an and
Beijing, respectively. During the reference periods the measured average
concentrations were 140 ± 99 µg m−3 in Xi'an and 75 ± 61 µg m−3 in Beijing. The relative composition of the
NR-PM2.5 evolved substantially during the extreme haze periods, with
increased contributions of the inorganic components (mostly sulfate and
nitrate). Our results suggest that the high relative humidity present during
the extreme haze events had a strong effect on the increase of sulfate mass
(via aqueous phase oxidation of sulfur dioxide). Another relevant
characteristic of the extreme haze is the size of the measured particles.
During the extreme haze events, the AMS showed much larger particles, with a
volume weighted mode at about 800 to 1000 nm, in contrast to about 400 nm
during reference periods. These large particle sizes made the use of the
PM2.5 inlet crucial, especially during the severe haze events, where 39 ± 5 % of the mass would have been lost in the conventional PM1
(particulate matter with aerodynamic diameter ≤ 1 µm) inlet. A
novel positive matrix factorization procedure was developed to apportion the
sources of organic aerosols (OA) based on their mass spectra using the
multilinear engine (ME-2) controlled via the source finder (SoFi). The
procedure allows for an effective exploration of the solution space, a more
objective selection of the best solution and an estimation of the rotational
uncertainties. Our results clearly show an increase of the oxygenated
organic aerosol (OOA) mass during extreme haze events. The contribution of
OOA to the total OA increased from the reference to the extreme haze periods
from 16.2 ± 1.1 to 31.3 ± 1.5 % in Xi'an and from 15.7 ± 0.7 to 25.0 ± 1.2 % in Beijing. By contrast, during
the reference periods the total OA mass was dominated by domestic emissions
of primary aerosols from biomass burning in Xi'an (42.2 ± 1.5 % of
OA) and coal combustion in Beijing (55.2 ± 1.6 % of OA). These two
sources are also mostly responsible for extremely high polycyclic aromatic
hydrocarbon (PAH) concentrations measured with the AMS (campaign average of
2.1 ± 2.0 µg m−3 and frequent peak concentrations above 10 µg m−3). To the best of our knowledge, this is the first data set
where the simultaneous extraction of these two primary sources could be
achieved in China by conducting on-line AMS measurements at two areas with
contrasted emission patterns.
Source apportionment analysis of hourly resolved particulate matter (PM) speciation data was performed using positive matrix factorization (PMF). The data were measured at an urban site in downtown Toronto, Canada during two campaign periods (April–July, 2013; November, 2013–February, 2014), and included trace metals, black carbon, and mass spectra for organic and inorganic species (PMFFull). The chemical composition was measured by collocated high time resolution instrumentation, including an Aerosol Chemical Speciation Monitor, an Xact metals monitor, and a seven-wavelength Aethalometer. Separate PMF analyses were conducted using the trace metal only data (PMFmetal) and organic mass spectra only (PMForg), and compared with the PMFFull results. Comparison of these three PMF analyses demonstrated that the full analysis offered many advantages in the apportionment of local and regional sources compared to using the organic or metals data individually. In combining the high time resolution data, this analysis enabled i) the quantification of metal-rich sources of PM2.5 (PM
Daily 24-hour PM2.5 samples were collected continuously from January 1 to December 31, 2010. Elemental concentrations from Al to Pb were obtained using particle induced X-ray emission (PIXE) method. This was the first full year continuous daily PM2.5 elemental composition dataset in Beijing. Source apportionment analysis was conducted on this dataset using the positive matrix factorization method. Seven sources and their contributions to the total PM2.5 mass were identified and quantified. These include secondary sulphur– 13.8 μg/m 3 , 26.5%; vehicle exhaust– 8.9 μg/m 3 , 17.1%; fossil fuel combustion– 8.3 μg/m 3 , 16%; road dust– 6.6 μg/m 3 , 12.7%; biomass burning– 5.8 μg/m 3 , 11.2%; soil dust– 5.4 μg/m 3 , 10.4%; and metal processing– 3.1 μg/m 3 , 6.0%. Fugitive dusts (including soil dust and road dust) showed the highest contribution of 20.7 μg/m 3 in the spring, doubling those in other seasons. On the contrary, contributions of the combustion source types (including biomass burning and fossil fuel combustion) were significantly higher in the fall (14.2 μg/m 3 ) and in the winter (24.5 μg/m 3 ) compared to those in the spring and summer (9.6 and 8.0 μg/m 3 , respectively). Secondary sulphur contributed the most in the summer while vehicle exhaust and metal processing sources did not show any clear seasonal pattern. The different seasonal highs and lows from different sources compensated each other. This explains the very small seasonal variations (< 20%) in the total PM2.5.
This paper compares and evaluates four principal methods of dust source and plume identification using MODIS data. The four MODIS methods used here are: (i) Roskovensky and Liou's dust identification algorithm, (ii) Ackerman's model, (iii) Normalized Difference Dust Index (NDDI), and (iv) Deep Blue algorithm. These techniques were applied to three recent significant events in the Middle East region. In addition, true color images and the HYbrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT) model were used to evaluate the result of each technique. To optimize the result of dust detection for each technique, the published dust–nondust thresholds had to be considerably adjusted on an event-by-event basis. Results show that techniques that use brightness temperature (BT) difference are the most reliable techniques for dust source detection in several situations such as multiplume and multimineralogical conditions (unlike the NDDI index and other optical based algorithms). However, these techniques cannot effectively differentiate dust plume from the bright desert surfaces (due to the same thermal behavior of dust and desert surfaces in both BT31 and BT32). This weakness impelled us to develop a new model based on Ackerman's technique because of its more precise results in dust source identification. In this new presented model called Middle East Dust Index (MEDI), BT29 was involved to highlight the difference between dust and desert surfaces as the [(BT31–BT29)/(BT32–BT29)] equation. In this equation, the values of dusty pixels are less than 0.6 while nondusty pixels are greater than 0.6. Results indicate that the MEDI model is ideal in both identifying dust plume and sources and desert surfaces. Finally, due to some misclassification of the MEDI model in differentiating cirrus clouds from dust plumes, the NDDI index was added to the initial model to distinguish them more accurately.
1] Chemical characteristics of inorganic water-soluble aerosol particles measured in large Asian springtime dust events during the Asian Pacific Regional Characterization Experiment (ACE-Asia) were investigated. Three specific flights (flights 6, 7, and 10) in the Yellow Sea boundary layer with high mineral dust concentrations mixed with pollutants from Asian urban centers are presented. Measurements during a similar campaign, Transport and Chemical Evolution over the Pacific (TRACE-P), in the same region suggested that fine-particle ammonium sulfate and nitrate salts, and potassium, apparently from biomass burning, are common particle ionic constituents in polluted air. Observations from the ACE campaign show similar characteristics and found that the main component of water-soluble mineral dust was Mg 2+ and Ca 2+ . Ion charge balances of measured fine and total aerosol suggest that a significant fraction of the Mg 2+ and Ca 2+ observed were in the form of carbonates. In polluted air mixed with dust that advected directly from large urban regions in roughly half a day to 1 day (flights 6 and 7), much of the fine-particle nitrate and sulfate (approximately 80%) was apparently associated with ammonium or potassium, the rest likely associated with mineral dust. Only air masses that spent 2–5 days over the Yellow Sea (flight 10) had clear evidence of Cl À depletion. Initial mass accommodation coefficients much less than 0.1 for uptake of SO 2 or HNO 3 by mineral dust in urban plumes containing fossil fuel and biomass-burning emissions could explain the observations. The data suggest an accommodation coefficient dependence on relative humidity.
This is an article from the journal, Remote Sensing of Environment [© Elsevier]. The definitive version is available at: http://www.sciencedirect.com/science/journal/00344257 The impact of mineral aerosol (dust) in the Earth's system depends on particle characteristics which are initially determined by the terrestrial sources from which the sediments are entrained. Remote sensing is an established method for the detection and mapping of dust events, and has recently been used to identify dust source locations with varying degrees of success. This paper compares and evaluates five principal methods, using MODIS Level 1B and MODIS Level 2 aerosol data, to: (a) differentiate dust (mineral aerosol) from non-dust, and (2) determine the extent to which they enable the source of the dust to be discerned. The five MODIS L1B methods used here are: (1) un-processed false colour composite (FCC), (2) brightness temperature difference, (3) Ackerman's (1997: J.Geophys. Res., 102, 17069–17080) procedure, (4) Miller's (2003:Geophys. Res. Lett. 30, 20, art.no.2071) dust enhancement algorithm and (5) Roskovensky and Liou's (2005: Geophys. Res. Lett. 32, L12809) dust differentiation algorithm; the aerosol product is MODIS Deep Blue (Hsu et al., 2004: IEEE Trans. Geosci. Rem. Sensing, 42, 557–569), which is optimised for use over bright surfaces (i.e. deserts). These are applied to four significant dust events from the Lake Eyre Basin, Australia. OMI AI was also examined for each event to provide an independent assessment of dust presence and plume location. All of the techniques were successful in detecting dust when compared to FCCs, but the most effective technique for source determination varied from event to event depending on factors such as cloud cover, dust plume mineralogy and surface reflectance. Significantly, to optimise dust detection using the MODIS L1B approaches, the recommended dust/non-dust thresholds had to be considerably adjusted on an event by event basis. MODIS L2 aerosol data retrievals were also found to vary in quality significantly between events; being affected in particular by cloud masking difficulties. In general, we find that OMI AI and MODIS AQUA L1B and L2 data are complementary; the former are ideal for initial dust detection, the latter can be used to both identify plumes and sources at high spatial resolution. Overall, approaches using brightness temperature difference (BT10–11) are the most consistently reliable technique for dust source identification in the Lake Eyre Basin. One reason for this is that this enclosed basin contains multiple dust sources with contrasting geochemical signatures. In this instance, BTD data are not affected significantly by perturbations in dust mineralogy. However, the other algorithms tested (including MODIS Deep Blue) were all influenced by ground surface reflectance or dust mineralogy; making it impossible to use one single MODIS L1B or L2 data type for all events (or even for a single multiple-plume event). There is, however, considerable potential to exploit this anomaly, and to use dust detection algorithms to obtain information about dust mineralogy. Accepted for publication
Data characterizing daily integrated particulate matter (PM) samples collected at the Jefferson Street monitoring site in Atlanta, GA, were analyzed through the application of a bilinear positive matrix factorization (PMF) model. A total of 662 samples and 26 variables were used for fine particle (particles < or = 2.5 microm in aerodynamic diameter) samples (PM2.5), and 685 samples and 15 variables were used for coarse particle (particles between 2.5 and 10 microm in aerodynamic diameter) samples (PM10-2.5). Measured PM mass concentrations and compositional data were used as independent variables. To obtain the quantitative contributions for each source, the factors were normalized using PMF-apportioned mass concentrations. For fine particle data, eight sources were identified: SO4(2-) -rich secondary aerosol (56%), motor vehicle (22%), wood smoke (11%), NO(3-) -rich secondary aerosol (7%), mixed source of cement kiln and organic carbon (OC) (2%), airborne soil (1%), metal recycling facility (0.5%), and mixed source of bus station and metal processing (0.3%). The SO4(2-) -rich and NO(3-) -rich secondary aerosols were associated with NH(4+). The SO4(2-) -rich secondary aerosols also included OC. For the coarse particle data, five sources contributed to the observed mass: airborne soil (60%), NO(3-)-rich secondary aerosol (16%), SO4(2-) -rich secondary aerosol (12%), cement kiln (11%), and metal recycling facility (1%). Conditional probability functions were computed using surface wind data and identified mass contributions from each source. The results of this analysis agreed well with the locations of known local point sources.
In March 2020, COVID-19 was officially classified as a pandemic and as a consequence people have adopted strenuous measures to prevent infection, such as the wearing of PPE and self-quarantining, with no knowledge of when the measures will no longer be necessary. Coronavirus has long been known to be non-infectious when airborne; however, studies are starting to show that the virus can infect through airborne transmission and can remain airborne for a significant period of time. In the present study, a spark-induced plasma spectroscopy was devised to characterize the air propagation of the virus in real-time. The risk of air propagation was evaluated in terms of changes in virus concentration with respect to distance traveled and measurement time. Thus, our study provides a benchmark for performing real-time detection of virus propagation and instantaneous monitoring of coronavirus in the air.
We demonstrate in this work online in situ characterization of potash fertilizer, a powder material, at its final production stage in factory on the production conveyer belt for quality assessment,...
The occurrence of atmospheric fine particles (PM2.5)-associated polycyclic aromatic hydrocarbons (PAHs), trace metals and organic molecular markers was investigated by conducting an intensive sampling campaign at the Eastern Mediterranean urban area of Nicosia (Cyprus). Sixty-two 24-hr PM2.5 samples were collected and analyzed for fifty parent and alkylated PAHs, twenty-five long chain n-alkanes, seventeen hopanes and twelve steranes used for source apportionment. The same number and kind of samples were analyzed to determine twenty-eight trace metals. Emphasis was given to investigate the air levels of the scarcely monitored although highly carcinogenic PAHs such as dibenzopyrenes, dibenzoanthracenes, 7H-benzo[c]fluorene and 5-methyl-chrysene, not included in the USEPA's sixteen PAH priority list (USEPA-16). UNMIX receptor model was applied to apportion the sources of atmospheric emissions of the determined organic compounds and trace metals and evaluate their daily contributions to the corresponding PM2.5 associated concentrations. For comparison purposes, principal component analysis with multiple linear regression (PCA/MLR) was also applied and its results are reported. The UNMIX receptor model, compared to PCA/MLR, offered a more precise source profile and more reliable daily mass source distributions by eliminating negative contributions. The individual and cumulative multi-pathway lifetime cancer risk (posed via inhalation, ingestion and dermal contact) by exposure to PM2.5-associated USEPA-16 listed and non-listed PAHs and selected airborne trace metals (As, Cd, Co, Ni, and Pb) were assessed. To estimate the contribution of each emission source to the total cancer risk, multiple linear regression analysis was performed, using as independent variables the daily source mass contributions and as dependent variables the respective cancer risk units. The estimated total cumulative cancer risk comprising all toxic PAHs, besides those included in the priority list, and metals was higher than the USEPA's threshold by a factor of eight, denoting a potential risk for long-term exposure of a population in the urban environment.
One major technical difficulty of laser-induced breakdown spectroscopy (LIBS) lies in achieving ideal accuracy of quantitative determination of the multiple chemical components in a target sample. In this study, we propose a LIBS multi-component quantitative analytical method based on the construction of a deep convolutional neural network (CNN). More than 1400 LIBS spectra, collected from 23 China national standard reference materials, were utilized to train the CNN and validate its predictive ability as well. The experiment was implemented by the LIBS system in MarSCoDe, which would be the Mars Surface Composition Detector on the rover of China's first Mars exploration mission in 2020. To evaluate the performance of the CNN, we inspect the root mean square error (RMSE) value of the prediction, with both overall RMSE and component-wise RMSE considered, and we further look into the prediction relative error of each component. We compare the performance of the CNN with that of two alternative schemes based on back-propagation neural network (BPNN) and partial least squares (PLS) regression respectively, with the PLS scheme actually containing two methods, i. e. PLS1 and PLS2. Besides the examination of the specific values of RMSE and relative error, we have also carried out some statistical analysis to endow the comparison with statistical significance. Moreover, we investigate the effect of baseline removal preprocessing upon the predictive ability of each method. The results show that the CNN method has the best performance among the four methods in terms of overall accuracy, no matter the test is based on the spectra with or without baseline removal, and the superiority of the CNN over the other three methods is more significant in the latter case. Since the number of samples is relatively small, the results demonstrated in this work are preliminary and unsuitable for immediate generalization, but they indicate that the CNN-based methodology is a promising tool for LIBS quantitative analysis with good accuracy and high efficiency.
Rapid and accurate identification of multiple types of rocks using spectroscopic techniques has a wide market application prospect and is always challenging due to similar chemical composition and complex matrix effects. In recent years, laser induced breakdown spectroscopy (LIBS) coupled with supervised machine learning and chemometrics methods (e.g. k-nearest neighbor (kNN), support vector machine (SVM), partial least squares (PLS), artificial neural network (ANN)) and combined with feature engineering techniques (e.g. principal component analysis (PCA)), has demonstrated great capabilities for efficient identification of materials with similar chemical composition. To further increase the classification accuracy, LIBS coupled with a convolutional neural network with two-dimensional input (2D CNN) is here investigated for the identification of rock samples, including dolomites, granites, limestones, mudstones and shales. A regularized network structure was first designed, according to the performance of validation dataset, to enable the most reliable discrimination of the rock specimens. The accuracy of test dataset was then evaluated by the determined model. Results indicated that validation and test set of the 2D CNN was able to reach an accuracy of 0.9877 and 1, respectively. Finally, the performance was compared with other identification methods, including: one-dimensional convolutional neural network (1D CNN), kNN, PCA-kNN, SVM, PCA-SVM, PLS-DA, and Human-Assisted ANN (HA-ANN). The proposed approach has demonstrated that CNN has a great potential for the lithological identification and could be a feasible and useful tool for LIBS spectral data processing.
Background
Yellow Dust (YD) is a natural source of particulate matter (PM) in Korea. It remarkably increases the concentration of PM. However, characteristics of PM in YD period are different from those of PM in non-YD period.
Objectives
To investigate whether the association of PM with mortality is different between all days and non-YD days in Seoul, Korea, 1998–2015.
Methods
We applied time-stratified case-crossover design to estimate effects of PM10 and PM2.5 on non-accidental cardiovascular and respiratory mortality. Effect estimates of PM were compared for all days in the study period and days without YD events. To identify whether different effect estimates between all days and non-YD days were not merely caused by the exclusion of high PM concentrations but rather by YD itself, we estimated effects of PM by randomly excluding the same number of days as days of YD.
Results
A total of 4,509,392 deaths were observed during the study period. A 10 μg/m³ increase in PM10 or PM2.5 was associated with a 0.15% (95% CI: 0.06% to 0.24%) or 0.27% (95% CI: 0.07% to 0.47%) increase in risk of non-accidental mortality for all days, respectively. These associations were changed to 0.30% (95% CI: 0.18% to 0.42%) and 0.33% (95% CI: 0.10% to 0.55%) when YD days were excluded from analyses. We also found that effect estimates of PM were larger when YD days were excluded than those when high PM concentrations were randomly excluded.
Conclusions
The effect estimates of PM differed between all days and non-YD days. Our study suggests that including YD days in the analyses is likely to attenuate the effect of PM in a usual urban environment.
Chemical analysis is commonly used in the field of forensic science where the precise discrimination of primary evidence is of significant importance. Laser-Induced Breakdown Spectroscopy (LIBS) exceeds other spectroscopic methods in terms of the time required for pre- and post-sample preparation, the insensitivity to sample phase state be it solid, liquid, or gas, and the detection of two-dimensional spectral mapping from real time point measurements. In this research, fingerprint samples on various surface materials are considered in the chemical detection and reconstruction of fingerprints using the two-dimensional LIBS technique. Strong and distinct intensities of specific wavelengths represent visible ink, natural secretion of sweat, and contaminants from the environment, all of which can be present in latent fingerprints. The particular aim of the work presented here is to enhance the precision of the two-dimensional recreation of the fingerprints present on metal, plastic, and artificially prepared soil surface using LIBS with principal component analysis. By applying a distinct wavelength discrimination for two overlapping fingerprint samples, separation into two non-identical chemical fingerprints was successfully performed.
Naturally-occurring Yellow Dust outbreaks, which are produced by winds flowing to Korea from China and Mongolia, create air pollution. Although there is a seasonal pattern of this phenomenon, there exists substantial variation in its timing, strength, and location from year to year. To warn residents about air pollution in general, and about these dust storms in particular, Korean authorities issue different types of public alerts. Using birth certificate data on more than 1.5 million babies born between 2003 and 2011, we investigate the impact of air pollution, and the avoidance behavior triggered by pollution alerts on various birth outcomes. We show that air pollution rises during Yellow Dust outbreaks and that exposure to air pollution during pregnancy has a significant negative impact on birth weight, the gestation weeks of the baby, and the propensity of the baby being born low weight. Public alerts about air quality during pregnancy help mitigate the adverse effect of pollution on fetal health. The results provide evidence for the effectiveness of pollution alert systems in promoting public health. They also underline the importance of taking into account individuals' avoidance behavior when estimating the impact of air quality on birth outcomes. We show that when the preventive effect of public health warnings is not accounted for, the estimated relationship between air pollution and infant health is reduced by more than fifty percent. In summary, air pollution has a deteriorating impact on newborns' health, and public alerts that warn individuals about increased air pollution help alleviate the negative impact.
The aim of the present study is to identify meat species by using laser-induced breakdown spectroscopy (LIBS). Elemental composition differences between meat species were used for meat identification. For this purpose, certain amounts of pork, beef and chicken were collected from different sources and prepared as pellet form for LIBS measurements. The obtained LIBS spectra were evaluated with some chemometric methods, and meat species were qualitatively discriminated with principal component analysis (PCA) method with 83.37% ratio. Pork-beef and chicken-beef meat mixtures were also analyzed with partial least square (PLS) method quantitatively. Determination coefficient (R(2)) and limit of detection (LOD) values were found as 0.994 and 4.4% for pork adulterated beef, and 0.999 and 2.0% for chicken adulterated beef, respectively. In the light of the findings, it was seen that LIBS can be a valuable tool for quality control measurements of meat as a routine method.
Deep neural nets with a large number of parameters are very powerful machine learning systems. However, overfitting is a serious problem in such networks. Large networks are also slow to use, making it difficult to deal with overfitting by combining the predictions of many different large neural nets at test time. Dropout is a technique for addressing this problem. The key idea is to randomly drop units (along with their connections) from the neural network during training. This prevents units from co-adapting too much. During training, dropout samples from an exponential number of different "thinned" networks. At test time, it is easy to approximate the effect of averaging the predictions of all these thinned networks by simply using a single unthinned network that has smaller weights. This significantly reduces overfitting and gives major improvements over other regularization methods. We show that dropout improves the performance of neural networks on supervised learning tasks in vision, speech recognition, document classification and computational biology, obtaining state-of-the-art results on many benchmark data sets. © 2014 Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever and Ruslan Salakhutdinov.
As the widespread application of online instruments penetrates the environmental fields, it is interesting to investigate the sources of fine particulate matter (PM2.5) based on the data monitored by online instruments. In this study, online analyzers with 1-h time resolution were employed to observe PM2.5 composition data, including carbon components, inorganic ions, heavy metals and gas pollutants, during a summer in Beijing. Chemical characteristics, temporal patterns and sources of PM2.5 are discussed. On the basis of hourly data, the mean concentration value of PM2.5 was 62.16±39.37μgm(-3) (ranging from 6.69 to 183.67μgm(-3)). The average concentrations of NO3(-), SO4(2-), NH4(+), OC and EC, the major chemical species, were 15.18±13.12, 14.80±14.53, 8.90±9.51, 9.32±4.16 and 3.08±1.43μgm(-3), respectively. The concentration of PM2.5 varied during the online-sampling period, initially increasing and then subsequently decreasing. Three factor analysis models, including principal component analysis (PCA), positive matrix factorization (PMF) and Multilinear Engine 2 (ME2), were applied to apportion the PM2.5 sources. Source apportionment results obtained by the three different models were in agreement. Four sources were identified in Beijing during the sampling campaign, including secondary sources (38-39%), crustal dust (17-22%), vehicle exhaust (25-28%) and coal combustion (15-16%). Similar source profiles and contributions of PM2.5 were derived from ME2 and PMF, indicating the results of the two models are reasonable. The finding provides information that could be exploited for regular air control strategies.
The large similarity existing in the spectral emissions collected from organic compounds by laser-induced breakdown spectroscopy (LIBS) is a limiting factor for the use of this technology in the real world. Specifically, among the most ambitious challenges of today's LIBS involves the recognition of an organic residue when neglected on the surface of an object of identical nature. Under these circumstances, the development of an efficient algorithm to disclose the minute differences within this highly complex spectral information is crucial for a realistic application of LIBS in countering explosive threats. An approach cemented on scatter plots of characteristic emission features has been developed to identify organic explosives when located on polymeric surfaces (teflon, nylon and polyethylene). By using selected spectral variables, the approach allows to design a concise classifier for alerting when one of four explosives (DNT, TNT, RDX and PETN) is present on the surface of the polymer. Ordinary products (butter, fuel oil, hand cream, olive oil and motor oil) cause no confusion in the decisions taken by the classifier. With rates of false negatives and false positives below 5%, results demonstrate that the classification algorithm enables to label residues according to their harmful nature in the most demanding scenario for a LIBS sensor.
The elemental composition of duplicate aerosol samples from north China, collected in particle size fractions by eight-stage cascade impactors on the Great Wall, near Beijing, is reported for 21 elements measured by particle induced X-ray emission analysis. Meteorological conditions during sampling on 1 April 1980 indicated that relatively clean northerly air from Mongolia and Siberia was sampled. Coarse terrestrial dust and additional fine aerosol components mainly of 0.5–1 μm aerodynamic diameter could be distinguished. Relative elemental abundances in the coarse mode resembled earth crust composition. Those in the fine mode resembled the South Pole aerosol, and fine mode elemental concentrations were low enough to suggest approximately background levels of several trace metals.
The authors have performed a structured expert judgement study of the population mortality effects of fine particulate matter (PM2.5) air pollution. The opinions of six European air pollution experts were elicited. The ability of each expert to probabilistically characterize uncertainty was evaluated using 12 calibration questions—relevant variables whose true values were unknown at the time of elicitation, but available at the time of analysis. The elicited opinions exhibited both uncertainty and disagreement. It emerged that there were significant differences in expert performance. Two combinations of the experts’ judgements were computed and evaluated—one in which each expert's views received equal weight; the other in which the expert's judgements were weighted by their performance on the calibration variables. When the performance of these combinations was evaluated the equal-weight combination exhibited acceptable performance, but was nonetheless inferior to the performance-based combination.
The partitioning of pollutant in the size-fractions of fine particles is particularly important to its migration and bioavailability in soil environment. However, the impact of pollution sources on the partitioning was seldom addressed in the previous studies. In this study, the method of continuous flow ultra-centrifugation was developed to separate three size fractions (<1μm, <0.6μm and <0.2μm) of the submicron particles from the soil polluted by wastewater and smelter dust respectively. The mineralogy and physicochemical properties of each size-fraction were characterized by X-ray diffraction, transmission electron microscope etc. Total content of the polluted metals and their chemical speciation were measured. A higher enrichment factor of the metals in the fractions of <1μm or less were observed in the soil contaminated by wastewater than by smelter dust. The organic substance in the wastewater and calcite from lime application were assumed to play an important role in the metal accumulation in the fine particles of the wastewater polluted soil. While the metal accumulation in the fine particles of the smelter dust polluted soil is mainly associated with Mn oxides. Cadmium speciation in both soils is dominated by dilute acid soluble form and lead speciation in the smelter dust polluted soil is dominated by reducible form in all particles. This implied that the polluted soils might be a high risk to human health and ecosystem due to the high bioaccessblity of the metals as well as the mobility of the fine particles in soil.
Combining the system of rapid collection of ambient particles and ion chromatography, the system of rapid collection of fine particles and ion chromatography (RCFP-IC) was established to automatically analyze on-line the concentrations of water-soluble ions in ambient particles. Here, the general scheme of RCFP-IC is described and its basic performance is tested. The detection limit of RCFP-IC for SO
42−, NO
3−, NO
2−, Cl− and F− is below 0.3 µg m−3. The collection efficiency of RCFP-IC increases rapidly with increasing sized particles. For particles larger than 300 nm, the collection efficiency approaches 100%. The precision of RCFP-IC is more than 90% over 28 repetitions. The response of RCFP-IC is very sensitive and no obvious cross-pollution is found during measurement. A comparison of RCFP-IC with an integrated filter measurement indicates that the measurement of RCFP-IC is comparable in both laboratory experiments and field observations. The results of the field experiment prove that RCFP-IC is an effective on-line monitoring system and is helpful in source apportionment and pollution episode monitoring.
The distribution patterns of the particulate matter (PM) and the associated elements were investigated from Seoul, Korea during spring 2001. The results of our measurements were analyzed to explain the behavior of metallic components by comparing their compositions mainly in terms of between Asian Dust (AD) and non-AD (NAD) period and between fine and coarse particle fraction. The computation of enrichment factor (EF) indicated that the magnitude of EF values for most hazardous metals during the AD period were even smaller than the NAD counterpart. The existence of low EF values during the AD period may be ascribable to the excessive input of crustal components like Al accompanied by the AD event. In accordance with this finding, the effects of the AD events were also reflected in diverse manners, when assessed by the concentration ratios of a given element for both AD/NAD period and fine-to-coarse (F/C) fraction. Results of this comparative analysis generally suggest that AD events are prominent sources for major crustal components in the fine particle fraction of PM. In addition, comparison of our measurement data with those obtained within the Korean peninsula and in the near-by Asian areas indicates that the metallic distribution patterns of the study area may be affected more sensitively by anthropogenic signatures. The results of our analysis, if investigated in relation with air mass movement patterns by means of the back-trajectory analysis, demonstrate consistently that the PM data measured during the study period can be closely tied with the signatures of both AD events and anthropogenic processes.
Restricted Boltzmann machines were developed using binary stochastic hidden units. These can be generalized by replacing each binary unit by an infinite number of copies that all have the same weights but have progressively more negative biases. The learning and inference rules for these “Stepped Sigmoid Units ” are unchanged. They can be approximated efficiently by noisy, rectified linear units. Compared with binary units, these units learn features that are better for object recognition on the NORB dataset and face verification on the Labeled Faces in the Wild dataset. Unlike binary units, rectified linear units preserve information about relative intensities as information travels through multiple layers of feature detectors. 1.