Michaël Rademaker

Correlation analysis of noise and ultrafine particle counts in a street canyon

Ultrafine particles (UFP, diameter<100 nm) are very likely to negatively affect human health, as underlined by some epidemiological studies. Unfortunately, further investigation and monitoring are hindered by the high cost involved in measuring these UFP. Therefore we investigated the possibility to correlate UFP counts with data coming from low-cost sensors, most notably noise sensors. Analyses are based on an experiment where UFP counts, noise levels, traffic counts, nitrogen oxide (NO, NO(2) and their combination NO(x)) concentrations, and meteorological data were collected simultaneously in a street canyon with a traffic intensity of 3200 vehicles/day, over a 3-week period during summer. Previous reports that NO(x) concentrations could be used as a proxy to UFP monitoring were verified in our setup. Traffic intensity or noise level data were found to correlate with UFP to a lesser degree than NO(x) did. This can be explained by the important influence of meteorological conditions (mainly wind and humidity), influencing UFP dynamics. Although correlations remain moderate, sound levels are more correlated to UFP in the 20-30 nm range. The particles in this size range have indeed rather short atmospheric residence times, and are thus more closely short-term traffic-related. Finally, the UFP estimates were significantly improved by grouping data with similar relative humidity and wind conditions. By doing this, we were able to devise noise indicators that correlate moderately with total particle counts, reaching a Spearman correlation of R=0.62. Prediction with noise indicators is even comparable to the more-expensive-to-measure NO(x) for the smallest UFP, showing the potential of using microphones to estimate UFP counts.

Predicting partial orders: ranking with abstention

The prediction of structured outputs in general and rankings in particular has attracted considerable attention in machine learning in recent years, and different types of ranking problems have already been studied. In this paper, we propose a generalization or, say, relaxation of the standard setting, allowing a model to make predictions in the form of partial instead of total orders. We interpret such kind of prediction as a ranking with partial abstention: If the model is not sufficiently certain regarding the relative order of two alternatives and, therefore, cannot reliably decide whether the former should precede the latter or the other way around, it may abstain from this decision and instead declare these alternatives as being incomparable. We propose a general approach to ranking with partial abstention as well as evaluation metrics for measuring the correctness and completeness of predictions. For two types of ranking problems, we show experimentally that this approach is able to achieve a reasonable trade-off between these two criteria.

A computational model of auditory attention for use in soundscape research

Urban soundscape design involves creating outdoor spaces that are pleasing to the ear. One way to achieve this goal is to add or accentuate sounds that are considered to be desired by most users of the space, such that the desired sounds mask undesired sounds, or at least distract attention away from undesired sounds. In view of removing the need for a listening panel to assess the effectiveness of such soundscape measures, the interest for new models and techniques is growing. In this paper, a model of auditory attention to environmental sound is presented, which balances computational complexity and biological plausibility. Once the model is trained for a particular location, it classifies the sounds that are present in the soundscape and simulates how a typical listener would switch attention over time between different sounds. The model provides an acoustic summary, giving the soundscape designer a quick overview of the typical sounds at a particular location, and allows assessment of the perceptual effect of introducing additional sounds.

Sampling approaches to predict urban street noise levels using fixed and temporary microphones

Requirements for static (prediction of L(den) and diurnal averaged noise pattern) and dynamic (prediction of 15 min and 60 min evolution of L(Aeq) and statistical levels L(A90,)L(A50) and L(A10)) noise level monitoring are investigated in this paper. Noise levels are measured for 72 consecutive days at 5 neighboring streets in an inner-city noise measurement network in Gent, Flanders, Belgium. We present a method to make predictions based on a fixed monitoring station, combined with short-term sampling at temporary stations. It is shown that relying on a fixed station improves the estimation of L(den) at other locations, and allows for the reduction of the number of samples needed and their duration; L(den) is estimated with an error that does not exceed 1.5 dB(A) to 3.4 dB(A) according to the location, for 90% of the 3 × 15 min samples. Also the diurnal averaged noise pattern can be estimated with a good accuracy in this way. It was shown that there is an optimal location for the fixed station which can be found by short-term measurements only. Short-term level predictions were shown to be more difficult; 7 day samples were needed to build models able to estimate the evolution of L(Aeq,60min) with a RMSE ranging between 1.4 dB(A) and 3.7 dB(A). These higher values can be explained by the very pronounced short-term variations appearing in typical streets, which are not correlated between locations. On the other hand, moderately accurate predictions can be achieved, even based on short-term sampling (a 3 × 15 minute sampling duration seems to be sufficient for many of the accuracy goals set related to static and dynamic monitoring). Finally, the method proposed also allows for the prediction of the evolution of statistical indicators.

Context-dependent environmental sound monitoring using SOM coupled with LEGION

Environmental sound measurement networks are increasingly applied for monitoring noise pollution in an urban context. Intelligent measurement nodes offer the opportunity to perform advanced analysis of environmental sound, but tradeoffs between cost and functionality still have to be made. When using a tiered architecture, local nodes with limited computing capabilities can be used to detect sound events of potential interest, which are then further analyzed by more powerful nodes. This paper presents a human-mimicking model for detecting rare and conspicuous sound events. Features encoding spectro-temporal irregularities are extracted from the sound, and a Self-Organizing Map (SOM) is used to identify co-occurring features, which most likely belong to a single sound object. Extensive training allows this map to be tuned to the typical sounds that are heard at the microphone location. A Locally Excitatory Globally Inhibitory Oscillator Network (LEGION) is used to group units of the SOM in order to construct distinct sound objects.

Prediction of ultrafine particle number concentrations in urban environments by means of Gaussian process regression based on measurements of oxides of nitrogen

Gaussian process regression is used to predict ultrafine particle (UFP) number concentrations. We infer their number concentrations based on the concentrations of NO, NO2, CO and O3 at half hour and 5?min resolution. Because UFP number concentrations follow from a dynamic process, we have used a non-stationary kernel based on the addition of a linear and a rational quadratic kernel. Simultaneous measurements of UFP and gaseous pollutants were carried out during one month at three sampling locations situated within a 1?km2 area in a Belgian city, Antwerp. The method proposed provides accurate predictions when using NO and NO2 as covariates and less accurate predictions when using CO and O3. We have also evaluated the models for different training periods and we have found that a training period of at least seven days is suitable to let the models learn the UFP number concentration dynamics in different typologies of traffic. Prediction of UFP number concentrations using Gaussian process regression.Simultaneous measurement at three urban sites of NO/NO2 and UFP.NO and NO2 are the inputs of the model; UFP is the target variable.Similar model performance at three urban sites at 5 and 30?min resolution.

Aggregation of monotone reciprocal relations with application to group decision making

This paper is composed of two complementary parts. The first part is a formal investigation into the interplay of properties of reciprocal relations, how monotonicity relates to some natural and intuitive properties, including stochastic transitivity. The goal is to aggregate monotone reciprocal relations on a given set of alternatives. Monotonicity is expressed w.r.t. a linear order on the set of alternatives. The second part is a practical protocol to both determine the best fitting linear order underlying the alternatives, and construct a reciprocal relation monotone w.r.t. it. We formulate the problem as an optimization problem, where the aggregated linear order is that for which the implied stochastic monotonicity conditions are closest to being satisfied by the distribution of the input monotone reciprocal relations. We show that if stochastic monotonicity conditions are satisfied, a monotone reciprocal relation is easily found on the basis of the (possibly constructed) stochastically monotone reciprocal distributional relation.

Noise measurements as proxies for traffic parameters in monitoring networks.

The present research describes how microphones could be used as proxies for traffic parameter measurements for the estimation of airborne pollutant emissions. We consider two distinct measurement campaigns of 7 and 12 days, at two different locations along the urban ring road in Antwerp, Belgium, where sound pressure levels and traffic parameters were measured simultaneously. Noise indicators are calculated and used to construct models to estimate traffic parameters. It is found that relying on different statistical levels and selecting specific sound frequencies permits an accurate estimation of traffic intensities and mean vehicle speeds, both for light and heavy vehicles. Estimations of R(2) values ranging between 0.81 and 0.92 are obtained, depending on the location and traffic parameters. Furthermore, the usefulness of these estimated traffic parameters in a monitoring strategy is assessed. Carbon monoxide, hydrocarbon and nitrogen oxide emissions are calculated with the airborne pollutant emission model Artemis. The Artemis outputs fed with directly measured and estimated traffic parameters (based on noise measurements) are very similar. Finally, a method is proposed to enable using a model calibrated at one location at another location without the need for new calibration, making it straightforward to include new measurement locations in a monitoring network.

Optimal monotone relabelling of partially non-monotone ordinal data

Noise in multi-criteria data sets can manifest itself as non-monotonicity. Work on the remediation of such non-monotonicity is rather scarce. Nevertheless, errors are often present in real-life data sets, and several monotone classification algorithms are unable to use such partially non-monotone data sets. Fortunately, as we will show here, it is possible to restore monotonicity in an optimal way, by relabelling part of the data set. By exploiting the properties of a (minimum) flow network, and identifying pleasing properties of some maximum cuts, an elegant single-pass optimal ordinal relabelling algorithm is formulated.

New Operations for Informative Combination of Two Partial Order Relations with Illustrations on Pollution Data

We discuss various ways in which to construct and process partial order relations or partially ordered sets (posets) in the context of ranking objects on the basis of multiple criteria. Oftentimes, it is undesirable or even impossible to devise a weighting scheme to compute a final score on the basis of the criteria. An alternative is then to restrict oneself to the information contained in the partial ordering of all objects implied by the criteria. We will consider some ways in which one can exploit partial order relations to determine a ranking of a collection of objects. More exactly, we will examine how to combine information coming from two sources, both for the case in which the sources are considered to be equally important, as well as for the case in which one source of information should take priority. We illustrate the concepts on pollution data coming from 59 regions in Baden-Württemberg.