[Interview] Gwendall Petit, France (Audio/GIS Engineer, Noise Mapping Project Manager)
Gwendall Petit is a French engineer in Geographic Information Science (GIS). He has been working in this field since 2008 and is a strong advocate of open science. He is an active member and one of the initiators of the Noise-Planet platform which aims to provide scientific tools for environmental noise assessment. This platform brings together researchers and engineers in acoustics, geographic information science and computer development. He is speaking here on behalf of this team.
NoiseCapture seems like a hybrid between foursquare and audio capture, what is it about?
NoiseCapture is a free and opensource application dedicated to the collaborative measurement of sound environments. Thanks to this tool, everyone is able to measure the sound levels around you, by associating the GPS track. Optionally, it is also possible to associate tags to characterize the sound environment and the level of pleasantness. Then, when measurements are uploaded to our server, we are feeding an online and interactive map (noise-planet.org/map_noisecapture/index.html). When you zoom in and click on a cell, you can have a lot of information about local sound levels and characteristics.
The reference to Foursquare is interesting. We tried to make the application as simple and fun as possible. Even if we didn’t have time to explore this more concretely, we would like to orient the application towards serious gaming, with sound quests and achievement levels.
Ok, and what is a NoiseCapture Party? It sounds like a party 🙂
In addition to this simple use, we have created the concept of NoiseCapture party (noise-planet.org/noisecapture_party.html). The idea is to organize an event on a given place and date, to invite people (members of an association, citizens, students, …) to collect noise levels in a more organized way. When measuring, the users just have to enter a specific code in the app so that all the contributions will feed, in near real time, a dedicated web map.
At the end of the party, we organize a debriefing with the participants to collect their feelings and remarks. It is a great way to initiate debates about noise pollution and then work on the implementation of solutions.
So far it works quite well, especially with teachers / researchers who use it with their students to make them aware of environmental noise issues but also with public authorities who want to initiate actions.
Who is running this project?
NoiseCapture is developed by a team of French public researchers and engineers, coming from Université Gustave Eiffel, CNRS and Cerema institutions (UMRAE and Lab-STICC laboratories: noise-planet.org/members.html).
This app is part of a wider project called Noise-Planet (noise-planet.org/), aiming at providing free and opensource tools to measure and simulate noise environments. In this context, we are also developing NoiseModelling (noise-planet.org/noisemodelling.html), a software allowing to produce noise maps.
Seems like a big project, when did it start?
We started working together in the team in 2008. The idea was originally to join acoustics and geographic information sciences into common research and tools. Indeed, at this period, there was a lack of tools able to mix these two approaches (e.g we were calculating a noise map (with in fact no base map layer) and then you had to export into a GIS to compute population exposure). We wanted to fill this gap by proposing an integrative approach.
In 2014 we had the opportunity to participate in the ENERGIC-OD project (cordis.europa.eu/project/id/620400), funded by the European Union. In the framework of this project we initiated the development of the NoiseCapture application in order to fill a gap we had identified: having a simple tool accessible to the largest number of people to measure noise levels in the environment, with as a background the principle of open science. The idea was also to take advantage of the large number of sensors available (smartphones) to have a large enough database to have something reliable, despite a lower theoretical quality.
At the end of the project in 2017, we released the app on the Google Play Store (play.google.com/store/apps/details?id=org.noise_planet.noisecapture ).
How many active users do you have?
Since the official release of the app in September 2017, NoiseCapture has 98 010 unique contributors, for a total of 418 685 measurements (tracks), corresponding to a total of 1 215 days of cumulative measurement.
We have around 12 000 active users, coming from all over the world. They send more or less 150 contributions per day.
The application almost lives by itself. It’s a success we didn’t expect at first. In our opinion, it proves that the issue of noise pollution is a subject that touches people.
Your work follows the open science principles, can you tell us more?
As researchers and engineers from French public organizations, we are committed to sharing our work with the wider community (researchers, companies, citizens, etc.).
It has become something almost normal nowadays, but when we initiated this approach, we were a bit alone.
We believe that opening up knowledge facilitates progress for the benefit of all. Moreover, this openness allows a better reproducibility of our work and thus ensures a better credibility of the results.
In these times, when there is constant talk of fake news and when there is a certain mistrust of science, this transparency seems important to us.
In concrete terms, we work for open science along these three axes:
- all our codes are freely available, through our Github repositories (e.g github.com/Universite-Gustave-Eiffel/NoiseCapture/)
- we open as much as possible the data we produce
- whenever possible, we publish our work in open journals (see noise-planet.org/publication.html)
In parallel, NoiseCapture allows us to include citizens in our scientific process. They become contributors and allow, at their level, to advance knowledge.
What about the data collected?
To make it simple, once the measurement is completed and described by the user, the contribution is sent to our server. The received file is then integrated into our system, after passing a control phase (e.g to remove outliers). From there, the data is made freely available, under ODbL licence, to the community through different channels (see noise-planet.org/map_noisecapture/noisecapture_exploit_data.html). One of the easiest ways to reuse this data is to use the .geojson file export that is done every night and provided by country and region (see data.noise-planet.org/noisecapture/) .
To date, the database is approximately 25 gigabytes in size. Without any pretension and from what we know, it is the largest open dataset in the world in this domain.
This amount of data must raise questions?
You are right. The reuse of these data is easy and free, but it nevertheless raises questions, especially on the aspect of quality. This is one of our research topics. We are particularly interested in the following points:
- How to characterize the quality of data from sensors (smartphones), having a lower acoustic quality than a class 1 microphone?
- How to “mix” these multiple and heterogeneous data with others from fixed and few sensors, but of good quality?
- How to manage temporal and spatial aspects? Is a measurement in a street at a given time representative of the sound environment at all times?
All these questions are not trivial. The NoiseCapture database allows us to work on them, with a volume large enough to allow us to better calibrate our methods.
Some publications to illustrate this point:
And what about the quality of the microphone?
Unlike the iPhone where there is only one sensor, which is well mastered, with Android we have a multitude of sensors, of different quality and range of measurements.
To try to have a better view on this point, we offer the possibility to calibrate the smartphone thanks to the application.
We propose different methods (noise-planet.org/noisecapture_calibration.html) but one of the simplest is to measure a noise in parallel with a reference sensor and then compare the decibels measured. Once the correction is applied, on the server side we know if the smartphone has been calibrated and, if so, with what gain.
In a technical point of view, to meet this need, we had to develop our own signal processing library.
Since you are recording the sound, what about privacy?
This point is one of the pillars of this project.
In NoiseCapture we are not recording the audio. We are “just” processing the sound to extract noise indicators (see : www.sciencedirect.com/science/article/pii/S2352340917303414?). This is a very important point because:
- It reduces a lot the amount of data to transfer to our server,
- We are sure to respect the privacy of our contributors, since it’s not possible to reconstruct the audio track and thus hear discussions.
At the same time, we collect very little user-related data:
- the user has a unique, randomly generated identifier. There is no notion of profile, account, password, … So we don’t know who corresponds to which identifier
- few information about the smartphone: brand, model, Android version. These data are used to work, a posteriori, on the data quality (e.g we can apply correction depending on the brand or microphone model).
Are you also available on iOS?
Unfortunately and up to now, NoiseCapture is available only on Android. We would like to develop also for the iPhone but it’s a huge task since we more or less have to redevelop the app from scratch (because the technologies and languages are not the same). It’s a big effort and currently we don’t have the money and human hands to do the job. We are open to any kind of help in this context!
What technologies did you use to write the Android version, and all the backend?
To develop NoiseCapture we used Android Studio and a set of free and opensource libraries such as JTransforms for the Fourrier transform (indicator calculations), MPAndroidChart for rendering charts or Supertooltips for GUI elements.
On the server side, our infrastructure is also based on free and opensource softwares. For example our database is stored and managed thanks to PostGreSQL / PostGIS and the map is rendered thanks to Geoserver and Leaflet.
What are the challenges or developments ahead?
One research question that is currently emerging concerns source recognition. As mentioned above, we do not have the audio track of the measurement. Therefore, how to recognize the sound source(s) (motorcycle, car, bird, …) using only the acoustic indicators. We are working with researchers in signal processing to develop automatic methods, notably based on classification by learning. Eventually, we would like to integrate this functionality in NoiseCapture to propose to automatically tag the measurement with the right sound source (the user would then only have to confirm).
A second line of research concerns the perception of soundscapes. Indeed, the link between “high decibel level” and “nuisance” is not systematic, and vice versa. It is thus interesting for us to associate the users’ feeling with the outcome of the measurement. This helps us to better characterize the measurement scene.
For this, we are working on improving the descriptors and pleasantness indicators, present in the description page, in order to follow the state of the art as closely as possible.