NETWORK OF LOW COST AUDIO SENSORS COUPLED WITH MACHINE LEARNING MODEL(S) TO AID IN PERSISTENT DETECTION AND IDENTIFICATION OF WILDLIFE SPECIES FOR BIODIVERSITY MONITORING

​

Description

We hypothesise that a confluence of low-cost consumer electronics (eg. USB microphones, Raspberry Pi computers, smartphones, lithium-ion batteries, small solar PV panels) together with open source Machine Learning (ML) models, make it possible to build an array of sensors for the persistent, automated detection and monitoring of wildlife, at significantly lower cost than has been possible to date using more manual, sporadic methods of measurement (eg. on-site surveys by ecologists).

​

With an increasing number of 'rewilding' and biodiversity monitoring projects ongoing worldwide, and an increasing range of attractive subsidy mechanisms coming into force to encourage such projects, we understand there to also be a growing need to monitor species present at a given location over time, so as to prove net gains in biodiversity eg. in order to receive subsidy payments, as well as to aid in wider species monitoring and conservation efforts.

 

At present much of this work is done manually by ecologists, but only for the limited time periods they are able to be on-site eg. with photographic or sound recording equipment, during which there is also inevitably some disturbance of an ecosystem through human presence.

 

An automated, persistent bioacoustic monitoring and identification system should aim to reduce cost (vs. repeated site visits) and dramatically increase data volume and quality. We also hope to prove that use of  machine learning models to sort signal from noise, de-duplicate and pre-categorise data can aid human ecologist in their analysis of monitoring data.

​

What we want to do

Establishment of a prototype device and small test network in a biodiverse area in a national park in the UK to see if the hypotheses above stand up to practical experimentation.

​

Anticipate a 3-month project from outset.

 

Outcomes expected to include:

• Determination of viability of producing such a network

• Determination of forecast cost to provide bioacoustic-detection-network-as-a-service

• Feedback from prospective clients for a network on functionality and pricing

​

Our unfair advantages

• 24/7 access to private plot of land in South Downs National Park with 25+ bird and mammal species already identified

• Access to pool of professional ecologists on weekly basis for advice and feedback

• Initial research completed to identify possible ML libraries, possible device design and component list, competing technologies at mature end (eg. radar based systems for bird detection at offshore wind farms) and lower end (smartphone apps, commercial audio capture devices used by ecologists in UK)

• Prospective reserved access to a cluster of low cost, specialised compute in the UK for ML model training

• Direct professional links to one of UK's foremost rewilding projects for feedback and possible later, commercial scale test

​

What we need

• £50,000 - MLOps Engineer time, Data Scientist time, hardware procurement, cloud services setup for data storage, budget for GPU-accelerated compute rental for ML model training

• We have all expertise and contacts within our existing network to complete this project