Monitoring wildlife in the Amsterdam Water Supply Dunes

Development of a cost-efficient automated wildlife camera network in a European Natura 2000 site

An innovative system for wildlife monitoring: autonomous cameras, automated data pipelines and species identification with AI

Managing and protecting nature requires effective and cost-efficient monitoring techniques. To achieve this, advanced technologies such as low-power digital sensors, wireless communication technology, and automated approaches are needed.

Researchers at the University of Amsterdam Institute for Biodiversity and Ecosystem Dynamics have teamed up with Waternet, the site manager at the Amsterdam Water Supply Dunes nature reserve, to accomplish the following aims:

- To develop and put into practice an innovative system for wildlife monitoring using an automated camera trap network

- To test the cost-efficiency of the automated system compared to manually applying traditional camera trapping

- To study the distribution, habitat use, activity, population structure and community composition of ground-dwelling mammals and birds

The Netherlands coastal landscape is dominated by sand dune ecosystems. The Amsterdam Water Supply Dunes, part of Kennemerland-Zuid Natura 2000 nature reserve, are used to purify two-thirds of the drinking water for the city of Amsterdam.

Nearby urbanization also makes it an important recreational area, with over 1 million people visiting annually for hiking and nature-oriented recreation. The first 100 meters from the shoreline are strictly managed for sea defence.

The study area is located within the Amsterdam Water Supply Dunes (AWD), a 34 km 2  dune ecosystem located west of Amsterdam and stretching 8 kilometers along the Dutch North Sea coast.

The AWD are owned by the Municipality of Amsterdam and managed by Waternet, a drinking water provider for Amsterdam and the surrounding area.

Dune habitats dominate the study area. Vegetation mainly consists of grasses (46%), but also includes large areas of scrublands (22%) and forests (21%), and smaller areas of sand (6%) and other low vegetation.

The landscape is important for a variety of vertebrates, vascular plants and invertebrates. Grazing mammals such as the European rabbit (Oryctolagus cuniculus) and the European fallow deer (Dama dama) are key species.

To examine the breeding success and winter mortality of rabbits, the site managers survey a 23.5 km long monitoring route each year in spring and autumn.

The surveys are done from a car driving along the monitoring route after sunset, and all rabbits that appear in the beam of the car light are counted.

A large population of fallow deer inhabits the nature reserve. In the winter of 2019-2020, the site managers installed 16 fenced exclosures ranging in size from 0.5–7.2 hectares to study how the dune ecosystem recovers after excluding high intensity grazing by fallow deer.

The fallow deer population is monitored every year in spring at sunset and sunrise within large sub-areas of the site. Counting is done from a car and multiple observers count at the same time.

Setting up a new wildlife camera network requires testing sensors under field conditions

The autonomous deployment of wireless 4G wildlife cameras with solar panels and automated data transmission was first tested in the years 2021–2023.

Setting up camera traps in the field

We used camera traps with wireless functionality like the one shown on the right for implementing the wildlife camera network. The wildlife cameras are triggered by a passive infrared sensor and use an infrared flash at night.

Because they are powered by solar panels and automatically transmit images (and a daily report) via 4G, this type of cameras allows for setting up an automated sensor network.

Cameras and solar panels were mounted with security cages on wooden poles, and heights between 30–50 cm were tested.

Viewing images, map locations and camera performance on the sensor portal

The autonomous deployment of wildlife cameras with solar panels did not require battery replacement or SD card retrieval during the pilot studies. Some deployments even lasted as long as 746 days (> 2 years).

Wireless data transmission also worked well, and scientists and site managers are able to view images and their camera locations via a web interface like the one shown on the right. The sensor portal also allows them to remotely monitor the performance of the cameras, including battery status, file uploads and potential malfunctioning.

Before installing the wildlife camera network, three pilots with 2–6 wildlife cameras each were conducted

The main objectives of the pilots were to test:

- The autonomous deployment of cameras with solar panels, no battery replacement and no SD card retrieval

- The detection of focal species like rabbit, fallow deer and red fox inside and outside fenced exclosures, and between regular and wide lens cameras

- Wireless data transmission with SIM cards and 4G via a telecommunication network and data accumulation of images over time

- Data accumulation of images over time

The most commonly detected species inside the fenced exclosures during the pilots were rabbits

Rabbits were mainly detected inside exclosures (pilots 1 and 2). Foxes were detected in low numbers in all pilot areas. Fallow deer were detected only outside exclosures (pilot 3).

For rabbits, more than 1,500 observations per year could be registered inside the three exclosures.

Besides rabbits, a total of 18 bird and mammal species were detected, and one toad.

The pilot study also paired cameras to compare those with a regular lens with those having a wide lens

While the total number of detections differed between lens angles, the detection rate of both species showed no statistically significant difference between wide and regular lens, as shown in the top figure. This suggests that species which frequently visit a site can be equally detected by both lens angles.

However, pilots 2 and 3 also showed that rare or less frequently visiting species such as mice, polecats, and various bird species often remain undetected by a regular lens, resulting in 2–6 times more species detected with a wide lens than with a regular lens, as shown in the bottom figure.

The final sampling design resulted in the implementation of a network with 65 cameras in July–November 2023

The sampling design of the camera network integrated 3 different design criteria:

- Grid-based sampling stratified by major habitats

- Along the rabbit monitoring route

- Paired sampling inside and outside fenced deer exclosures

This combined approach means that some cameras are used for 2 criteria.

For the grid-based sampling, we placed a 1 × 1 km 2   grid over the study area and located at least one camera per grid cell, which means that the sampling design evenly covers the site.

To specify the exact location of each camera, the camera placement was stratified by habitat type. The camera locations shown on the map are colored to indicate the habitat in which they have been placed.

This sampling approach can be used to study the occupancy and distribution of single species, but also the diversity and composition of the whole wildlife community. This is an expansion on the current monitoring which is mainly focused on rabbits and fallow deer.

Along the 23.5 km traditional rabbit monitoring route, cameras were placed within a buffer up to 20 meters from the route, taking locations of grid camera placements into account.

Cameras were placed approximately one kilometer apart.

For the fenced exclosures, a paired sampling design was used, with one camera placed inside each of the 16 exclosures and one control camera placed outside each exclosure.

Exclosure control cameras were placed within a buffer 200–300 m away from the exclosure boundary.

Exclosure cameras and their control cameras were placed within the same habitat type.

Placement of all 65 cameras in the network was stratified by major habitat type using a digital vegetation map of the study area.

In this way, the number of cameras was placed proportionally to the percentage area of each habitat type.

End-to-end data pipeline demonstrates the feasibility of automated biodiversity monitoring

The development of this wireless 4G wildlife camera network in the coastal dunes of the Netherlands provides an innovative example of how advanced technology can automate biodiversity monitoring.

Automated data transmission results in near-real-time data from wildlife cameras

The autonomous deployment of the cameras with solar panels and the wireless data transmission via 4G allows scientists and site managers via a web interface to view images in near-real time. A key advantage is that camera performance can be remotely monitored. Researchers can quickly pick up issues caused by low batteries, damage from wildlife, water ingress, or vegetation in front of cameras.

This can strongly reduce the staff costs for annual data collection because it requires less field visits than with traditional camera traps. It also allows site managers to rapidly share interesting wildlife observations among them or with visitors.

Processing workflows with AI are becoming crucial tools for the efficient and secure handling of wildlife camera images

Automated species identification was tested with a deep learning model from Conservation AI (https://www.conservationai.co.uk/). This detects and classifies species from camera trap images and was specifically fine-tuned to support the focal wildlife species (rabbits, foxes and fallow deer).

The deep learning model performed reasonably well for automatically identifying the three focal species, but requires further improvement.

The fox showed the highest balanced accuracy and recall of the three focal species, followed by the rabbit and the fallow deer.

The model performed particularly well when detecting humans, but not when detecting blank images.

Automated wildlife camera networks can be more cost-efficient than data collection and processing with traditional camera traps

To test the cost-efficiency of the system, the full economic costs of the automated camera network were calculated, including establishment and annual operation costs over a 5- and 10-year time period. These costs were compared to the costs of a manual system with traditional camera traps. Manual systems require replacement of batteries and SD cards, and species identification of images by humans, which can be very time consuming.

The total costs of the automated system over a 10-year period are 43% lower than a traditional system

We calculated an overall cost savings of over 500,000 EUR for the automated camera network over a 10-year period.

Over a 5-year period, overall costs can be over 230,000 EUR cheaper. The cost-efficiency of the automated camera network over a 5- and 10-year period remains nearly the same (-40% vs. -43%).

Establishment costs of the automated system are higher than traditional camera trapping because of the higher material costs. However, since the total annual costs are much lower, the automated system results in an overall cost saving a 5- or 10-year period.

The results demonstrate that wildlife monitoring can be operationalized with autonomous wildlife cameras, automated data pipelines and AI species identification using deep learning.

The system can be applied in other nature reserves in open habitats with mobile network coverage. Costs will be lower in countries with lower wages than the Netherlands. The availability of open-source wildlife and object detection models will further decrease the monitoring costs with wildlife cameras.

While the initial establishment costs can be higher, this pays off because automated systems can be more cost-efficient than the manual data collection and data processing with traditional camera traps.

Team

Science is not possible without ongoing collaboration between researchers and partners. The following researchers and organizations contributed substantially to this study: