SYnchronized Low power Versatile Acoustic Network Including Artificial intelligence – SylvanIA

Despite ambitious targets at reducing biodiversity loss, assessing these changes remains an epic challenge. A better monitoring of ecosystems remains the only solution. Acoustic passive monitoring is getting popular to sample different groups of vocal wildlife. However these tools remain limited as hardware is expensive and does not give a clear view of species at a large scale. This is due to: lack of accuracy in sound source positioning and tracking ; impossibility to improve it and to cover large areas by distributing sound antennas on several sensors ; lack of global ultra-precise synchronization of the sensor network. The storage and power limitations, high sampling frequency recordings are necessary for species detection, are crucial. Therefore, we propose from advanced AI to low cost embedded AI advances and precise localisation to allow relevant target recording and supply longer autonomy and accuracy of biodiversity surveys.
In SylvanIA (SYnchronized Low power Versatile Acoustic Network with embedded AI), we will develop a novel biodiversity sensing network paradigm for monitoring and tracking accurately biodiversity species on large survey areas. For that, we will join the forces of acoustic researchers and engineers, hardware and software development, and quantitative ecologists and conservation biologists, including a collaboration with the MFFP department of the ministry of Canada.
We aim to build a low cost, low power and precisely auto-synchronized (innovation 1, by radio and acoustic emissions from each node), distributed intelligent sensor network (innovation 2). Each node will have a large frequency band and versatile intelligent triggers (innovation 3) by joint research between IM2NP, LIS, and LEHNA and MFFP.
SylvaniA innovate AI solutions for such network. The recent deep learning advances there has been an exponential growth in the use of Deep Networks (DNs) on various time-series, with some promising results in bioacoustical transients (Ferrari Glotin 2021). However, the vast majority of DNs do not directly observe the time-series data but instead a handcrafted representation : vast majority of state-of-the-art methods combine DNs with some variant of a Time-Frequency Representation (TFR) an image representation of a time series, such as wavelets or localized complex sinusoids. It is for example common to employ wavelet transforms on biological signals and spectrogram on voices. These different TFR have different precisions. Hence, the choice of TFR has the potential to dim, or amplify, the precise bioacoustical content. In SylvaniA will take advantage of synchronous observation to adapt by Wigner Ville learnable decomposition in order to discriminate the different sources in time and space and frequency, and thus optimize the information content of the AI representations of the target species (innovation 4). Then TRL 6 pilot studies will be conducted with novel protocols (innovation 5), deployed in the Alpine area for studying specific bird species very sensitive to climate change, on borders of agricultural areas, humid zones, and on the Québec Biodiversity monitoring Network recently initiated to measure changes across ecosystems and communities in view of climate change.
A spatial and temporal map of distribution of acoustic diversity is expected as a result of this study, allowing to investigate movement and phenology strategies at both large and small spatio-temporal scales. Detection will be included into stochastic and imperfect detection models for better understanding how climate change and management practices can transform the sound landscape. Integrated in Québec and the Alpin monitoring network, proper result dissemination and knowledge transfer toward landscape managers and locales policies will be ensured.