Abstract
An important problem in forensic analyses is identifying the provenance of materials at a crime scene, such as biological material on a piece of clothing. This procedure, known as geolocation, is conventionally guided by expert knowledge of the biological evidence and therefore tends to be application-specific, labor-intensive, and subjective. Purely data-driven methods have yet to be fully realized due in part to the lack of a sufficiently rich data source. However, high-throughput sequencing technologies are able to identify tens of thousands of microbial taxa using DNA recovered from a single swab collected from nearly any object or surface. We present a new algorithm for geolocation that aggregates over an ensemble of deep neural network classifiers trained on randomly-generated Voronoi partitions of a spatial domain. We apply the algorithm to fungi present in each of 1300 dust samples collected across the continental United States and then to a global dataset of dust samples from 28 countries. Our algorithm makes remarkably good point predictions with more than half of the geolocation errors under 100 kilometers for the continental analysis and nearly 90% classification accuracy of a sample's country of origin for the global analysis. We suggest that the effectiveness of this model sets the stage for a new, quantitative approach to forensic geolocation.
Introduction
In forensic science, determining the "provenance" of evidence—such as a suspect's clothing or a shipping container—is a critical task. Traditionally, forensic geolocation has relied on the painstaking identification of microscopic pollen grains or soil minerals by specialized experts. While effective, this process is inherently subjective, slow, and difficult to scale.
The advent of high-throughput DNA sequencing offers a transformative alternative. It is now possible to identify tens of thousands of microbial species (bacteria, fungi, etc.) from a single swab. However, converting this high-dimensional "biological fingerprint" into precise geographic coordinates presents a significant computational challenge. This paper introduces DeepSpace, a novel algorithm designed to move forensic geolocation from subjective expert opinion to a purely data-driven, quantitative science.
Methodology: The DeepSpace Algorithm
DeepSpace bridges spatial statistics and deep learning through an ensemble approach. Rather than treating geolocation as a standard regression problem—which can be difficult for neural networks to optimize directly—the algorithm redefines it as a series of classification tasks.
- Spatial Partitioning: The map is randomly partitioned into discrete "tiles" using Voronoi diagrams.
- Deep Learning Classifiers: For each random partition, a Deep Neural Network (DNN) is trained to classify which tile a sample belongs to based on its fungal composition (the presence or absence of thousands of species).
- Ensembling: This process is repeated across $J=50$ different random partitions.
- Intensity Surface Generation: By averaging the results across these diverse "views" of the map, the algorithm produces a smooth, continuous probability surface. The peak of this surface serves as the predicted location.
This approach allows the model to capture complex, non-linear regional patterns while avoiding the "edge effects" or artificial boundaries typically associated with dividing a map into fixed zones.
Data and Experimental Design
The study evaluates the model across three distinct geographic scales:
- Continental U.S.: 1,301 dust samples collected from outdoor door trims by citizen scientists. This dataset featured a massive 57,331 distinct fungal phylotypes.
- Global: A dataset of 399 samples from 28 countries across six continents, containing 15,475 fungal phylotypes.
- Regional: A subset of 116 samples from three counties in North Carolina to test the limits of local resolution.
Key Insights
Fungi as Superior Indicators
The researchers found that fungal data is generally more informative for geolocation than bacterial data. This is likely because fungal spores are more diverse and geographically restricted compared to bacteria, providing a more distinct regional signature.
The Power of the Ensemble
DeepSpace significantly outperformed standard DNNs trained for simple state-level classification. By operating at the point level through random partitions, the model captures subtle regional patterns that a simple state-by-state classifier misses. It also outperformed previous methods like Bayesian Discriminant Analysis (BDA), which had a median error of ~263 km on the same U.S. dataset.
Resolution Limits
The regional analysis in North Carolina revealed a "resolution floor." While the model is highly effective at a national scale, it struggled to distinguish between adjacent counties, achieving only ~53% accuracy. This suggests that at a certain point, local environmental stochasticity outweighs broad biogeographical patterns.
Results
- National Accuracy: In the continental U.S., DeepSpace achieved a median absolute prediction error of 86.9 km. More than half of the predictions were within 100 km of the true origin.
- Global Accuracy: The model achieved 89.5% accuracy in identifying the correct country of origin. It was particularly adept at distinguishing between distant countries, though it occasionally faced challenges with bordering nations sharing similar climates.
Future Directions
While DeepSpace sets a new benchmark for quantitative forensics, several hurdles remain before it can be utilized in a courtroom. Future work must address seasonality, as fungal communities change throughout the year. Additionally, researchers aim to develop mixture models to reconstruct the travel history of objects that have moved through multiple locations, and integrate multiple biological kingdoms (fungi, bacteria, and pollen) into a single, unified classifier.
Summary updated