GEOGRAPHIC PROFILING
My core research investigates using geographic profiling (GP) methods to identify wildlife breeding sites from sighting data and assessing their efficacy across species and landscapes.
By assessing the performance of current Bayesian GP models using simulated data, my aim is to find the breaking point at which the models' accuracy fails when applied to real-world ecological data. This analysis will then act as a springboard for further research into developing GP models specially for ecological data analysis.
SPATIAL MODELS → ECOLOGICAL MODELS
01 // What is Geographic Profiling?
Theoretical Background
Geographic profiling is a technique developed in Criminology by Dr Kim Rossmo to help deal with the information overload associated with serial crimes. It allows law enforcement to more accurately focus investigation efforts by estimating a criminal's anchor point (home, workplace, etc.).
The method is based on a few key concepts:
- Principle of Least Effort - Excluding other external variables, an individual will choose to commit an act at the minimum distance required to reach a target
- Buffer Zone - The area immediately surrounding the anchor point has a low probability of incident due to the avoidance of “fouling the nest”, rather than a lack of ability or opportunity.
- Decay Distance- The probability of an incident decreases with distance from the anchor point.
Ultimately, these three principles intersect to form a predictive mathematical surface that transforms scattered incident data into a highly targeted search area.
02 // Geographic Profiling Showcase
ROSSMO_GP_MODEL
Grid-based profiling utilising Dr. Kim Rossmo's complete formula. Click to add incident points. Confidence is calculated via inverse spatial variance (Hit Score Area).
OBSERVATIONS
0
SPATIAL_CERTAINTY
0%
Awaiting telemetry input...
The Buffer Zone (B)
Notice the hollow "donut hole" at the centre of the probability cluster. This represents the biological buffer zone. Central place foragers rarely hunt immediately outside their dens to avoid attracting predators to their vulnerable young.
Distance Decay (f & g)
These parameters dictate the steepness of the probability drop-off. Parameter f controls how quickly probability fades outside the buffer, while g controls the sharp decline inside the buffer zone.
Spatial Certainty
Rather than a simple sample size metric, certainty is calculated using the inverse spatial variance (Hit Score Area). If the high-probability "hotzone" is smeared across the entire map, confidence is low. A sharp, dense mathematical peak results in high confidence.
03 // Bridging the ecological gap
Turning crime fighting methods into conservation tools
The same principles which make GP such a powerful tool for spatial analysis in criminology can also be applied to ecology. Most obviously, in the case of central place foragers (CPF). These CPF are species which are tethered to a specific anchor point to which it must return. Most commonly, these anchor points are the locations of roosts or nest sites, but they can also be linked to breeding sites and food caches.
The spatial ecology of CPF species is altered due to the constraint of being tethered to a specific anchor point, and as such their foraging patterns can be considered comparatively with Rossmo's original concepts.
- Principle of Least Effort - Excluding other external variables, an animal will choose to forage at the minimum distance required to reach a suitable habitat
- Buffer Zone - The area immediately surrounding the anchor point has a lower probability of foraging due to the avoidance alerting predators as to their nest location. Particularly poignant in the case of mothers with young offspring.
- Decay Distance- The probability of foraging decreases with distance from the anchor point. In an ecological context, Marginal Value Theorem states that an individual must balance the energetic cost of travel with the resource density of a foraging patch. In the case of single-load foragers such as raptors, this dictates that as distance from the nest increases, the quality of the prey must increase to justify the return journey.
04 // Agent_Based_Modeling
HOW CAN WE TEST THE BREAKING POINT OF GP MODELS?
While GP models are powerful, and have been used successfully in ecological applications, current methods have their limitations, they are strictly spatial tools which do not see the ecological big picture.
Animals do not move arbitrarily through their landscape, rather there are a multitude of factors which impact their behaviour. These can vary from avoidance of anthropogenic factors, to being drawn to specific habitats types to feed. Therefore, where current GP models may predict an anchor point may be skewed as a result of sighting density in favoured habitats, rather than considering the spatial distribution of resources in their priors.
In order to test the robustness of these models, we must first have sightings data across multiple species, landscapes and sampling methods. This will allow us to evaluate how well the models generalise to unseen data. However, in reality such datasets are often limited by the availability of data and the cost of collecting it. In order to overcome these limitations, and to allow us to replicate movement consistently, we are using an agent-based modelling approach to generate synthetic movement tracks from which we can take our samples.
05 // SIMULATING MOVEMENT
Animal Movement Simulator
Pure Brownian vs. Correlated Random Walk
Pure Brownian
Random direction at every step.
Correlated Random Walk
State: Standard Foraging (Avian Search)
The Brownian Baseline
Pure Brownian motion assumes each step is entirely independent. This results in erratic, highly localised thrashing with zero directional persistence. It fails to replicate the intentional, forward-moving nature of a living organism navigating its environment.
Correlated Trajectories (CRW)
A Correlated Random Walk (CRW) introduces mathematical memory. By constraining turning angles relative to the agent's current heading, the simulation produces the natural, sweeping trajectories and persistent momentum characteristic of actual movement behaviour.
Ecological Accuracy
This mathematical framework allows for complex, real-world modelling. By dynamically altering the movement correlation based on underlying raster data, the agent seamlessly transitions from rapid, straight-line travel into intense, area-restricted search (ARS) when entering high-quality habitat.
06 // expanding the movement model
A movement mechanism alone is not sufficient to generate realistic tracks for animals, we must consider what ecological factors influence movement.
Utilising the Levy flight model to simulate realistic movement paths provides us with a great foundation off which to define movement models via a step selection function. By altering the persistence degree of the agent based on its current state we can change how it moves when hunting vs returning to its anchor point vs exploring new areas.
The step selection function implementation also allows us to determine when and how the agent should change its movement strategy, this leads into a dynamic movement model wherein the agent interacts with a stack of environment rasters to determine its next step.
My current work is focussed around integrating these ecological rasters with the agent's movement model, allowing for avoidance of ecological obstacles, and a greater affinity for suitable habitat when in a foraging state.