An Indian, a German and a Dutch woman went on a mission: deploying cameras in the forest. What they didn’t know was that as soon as they crossed a barrier and were driving on the unpaved forest roads, they would get questioned a lot. Because what were three largely foreign women doing in the middle of the forest with a car? They don’t look like foresters, nor like hunters, they must be lost or…?
The responses to the sight of us are in fact polarising. While some are kindly offering their help to manoeuvre us out of the forest back onto the road, others perceive us as disturbance to wildlife or their property, not hesitating to take our pictures and threaten to take legal or less official measures to drive us out.
It also makes us wonder if the reactions were the same if we were a team of men. The fact that is still of seldom luck to ever meet another female employee in the forest stresses that women are still underrepresented in the field. At meetings where we introduce our project, a reoccurring question posed by the often entirely male attendees is “And who is going to do the field work?” Which we learnt to answer with confidence and a faint smile which recognizes but similarly refutes the male chauvinism behind the question.
“Us, of course”
But people are not the only challenge we, the “Deer Observers”, encounter in the field…
Early morning, we drive off with our rental car to the study area. Behind the gates, we encounter a network of rather adventurous forest roads that do not always deserve to be called such. Sudden dead ends on narrow roads are not uncommon, as are roads only existing in the memories of our navigation systems. Can this soft soil hold our weight? Is the ground clearance of our vehicle really high enough to survive this overgrown path unscathed? How do we get across this waterlogged pit? Driving in the forest sometimes means taking a risk or two. Keep calm and breathe…or reverse back out.
Eventually, we arrive at the first deployment site. Sometimes we get lucky, and our sample point is near the road, but every now and then it is steeped up, the path is muddy, waterlogged or turns out to be in a jungle of thorny underbrush. Once we get close (within 100 meters) to the sample point, we choose a suitable tree. That is, a tree sturdy enough to bear a camera but still within the diameter of the lock and with a good view towards the North, preferably at least 12 unobstructed meters. As you can imagine, this is not always a simple task when you find yourself in the vast open of a salvage-logged spruce stand.
Yet, once the best option is found, we start the deployment ritual. While one of us checks the camera settings and attaches the camera to the tree, the second team member starts a survey, where we e.g. save the coordinates and answer questions about the surroundings, such as the habitat type. In the meantime, the third person sets out, marking certain distances to take images of the canopy closure and vegetation cover for the so-called deployment calibration. About 15 minutes later we are done and continue to the next location to do this all over again.
Every month or so we return to replace batteries, clear upgrowing vegetation and exchange SD cards. Factoring in the challenges we come across, such as fallen trees blocking roads, extremely steep slopes, thunderstorms, we manage to deploy and/or check around 10 cameras daily.
And what’s next? Well, catching up with sleep and the weekend, obviously.
And then, once rested, it is our priority to save the data collected in the field. That is, thousands of images on the SD cards but also additional data describing the habitats surrounding our cameras, like canopy closure and vegetation cover. To give you an idea: in the first camera check we collected 715,798 images (499 GB), which excludes 37% of our total camera count (because other cameras were installed a bit later).
And then? Do we have to go through each image separately? The short answer is yes, but we receive support from dedicated camera trap management platforms engaging smart helpers – artificial intelligence. Specifically, we are working with an online server called “Agouti”, like the American rodent. It creates sequences (colliding all images taken without an interruption of at least 2 minutes) and identifies species. However, it sometimes identifies humans as cattle and deer as wolves.
Why Deer Observer?
As our prime interest is relieving the “deer-forestry conflict” which describes the concern of foresters and other stakeholders that deer have become overabundant in large parts of Europe and are now impeding the urgently needed forest transformation and reforestation by selectively browsing climate-resilient tree species. Insufficient regeneration of such species may entail impacts on forest biodiversity, ecological functioning and economic loss.
The reason for this perceived overabundance can be found in the lack of natural predators and centuries of misguided hunting as well as a specific type of forest management. However, not all stakeholders agree with the existence of such a man-made problem or see the need for a change in management strategies. The Hegejagd (“stag hunt”) traditionally practiced in Germany, for example, focuses on trophy hunting and seeks to maintain or increase deer densities for hunting ease and spruce is still the preferred species in the timber industry.
The data we collect shall provide unbiased estimates of the true population density as common ground to resolve conflicting perceptions and interests. Getting from cute images of deer and their offspring to population density estimates is of course not too straightforward. How do we know if two images of the same species represent the same or different individuals? The answer is, courtesy to some smart scientists like Marcus Rowcliffe, who came up with accurate statistical models, like the Random Encounter Model, we don’t need to know. The REM is based on the assumption that randomly placed camera traps will encounter randomly moving deer at rates that depend on the deer density and movement speed. Envision a playground with stationary objects like walls and climbing equipment and moving objects like balls being kicked around. The more balls there are within the area of the pit (ball density) and the faster they move, the more frequently they will collide with any stationary object. Hence, if we measure the speed of the balls and the number of collisions we can quite safely deduce ball density. Here, the balls are deer, and the stationary objects are camera traps.
Now, in this real-life example, encounter rates are more complex due to vegetation obstructing camera views. Smaller deer, like roe deer, are detected at shorter distances than larger species, like red deer. To address this, we estimate the detection zone separately for each camera deployment and species. This is done by analyzing the angles and distances at which each species is first detected. By linking pixel data from images to object size and position, and calibrating these with known object sizes at each site (using a method we call the pole-dance Picture 1), we can automatically determine these parameters when we mark the animals’ hoof positions.
In the Agouti programme, we can track the animal’s position over the period they are present within the detection zone and then easily estimate movement speed (Picture 2). This gives us all we need to estimate density (see Wearn et al. 2021 for a formal description).
The project work may be challenging on many levels, but it is fun and stimulating to work on these diverse tasks. Stay tuned for the first results from our research – and more adventures from our Deer Observer team.
