Community Memory Meets Machine Learning to Map Poaching Hotspots for Smarter IoT Deployment

…

…

The problem of deciding where to place limited monitoring infrastructure in a large conservation landscape is, at its mathematical core, a coverage optimization problem. When the number of possible sensor locations vastly exceeds the number of sensors that can be affordably deployed and maintained, the search space grows combinatorially, which is why the general formulation of sensor placement over large regions is classified as NP-hard. In practical terms, this means that no algorithm can guarantee finding the single best arrangement of devices across a sprawling landscape in a reasonable amount of time. Metaheuristic techniques such as genetic algorithms, particle swarm optimization, and ant-colony optimization have become the standard workarounds because they can produce good approximate solutions, but they share a critical dependency: they all require some prior assumption about where monitoring matters most. Without a spatial prior that ranks parts of the landscape by the likelihood of the events being monitored, these optimizers are essentially searching blind.

This dependency explains why the cold-start situation described in the study is so consequential. Established anti-poaching analytics platforms, including PAWS and systems built on SMART patrol protocols, have repeatedly shown that wildlife crime is not randomly scattered across a protected area. Instead, illegal activity concentrates in predictable ways relative to environmental and human-made features. Poachers tend to favor locations that offer concealment, quick access, and low probability of encounter with rangers, and they avoid places with heavy human presence or open sightlines. Machine-learning models trained on years of patrol records can capture these regularities and convert them into risk maps that guide patrol planning. But every one of these systems begins its lifecycle with a dataset of past incidents, and that is precisely the asset a cold-start site does not have.

The University of Hyderabad campus illustrates how common this gap is. Although the site spans roughly 2,000 acres and supports diverse mammal, bird, and reptile populations, there are no formal incident logs, no structured ranger patrols, and no curated enforcement database documenting where poaching has occurred. What does exist is a different kind of record: the accumulated, informal memory of security staff, campus workers, field researchers, and long-term residents who have observed or heard about poaching activity over many years. This kind of distributed knowledge is widespread in conservation settings across developing regions, but also in urban forests, fragmented habitats, and community-managed reserves worldwide. The methodological challenge is that such knowledge is anecdotal, unverified, and not machine-readable, so it cannot be fed directly into the spatial models that conventional wildlife-crime prediction relies upon.

The five-stage pipeline presented in the study can be understood as a systematic translation layer between human memory and geospatial data science. The first stage establishes a digital skeleton of the landscape using OpenStreetMap and QGIS, which are open tools that make the approach reproducible by other institutions without licensing costs. Digitizing a land parcel from scratch is a nontrivial step, because the resulting vector layers of natural features such as water bodies, vegetation patches, and rocky outcrops, together with infrastructural elements like roads, buildings, and walls, become the reference geometry against which all subsequent measurements are made. The second stage enriches this skeleton with satellite-derived information drawn from four complementary sources: the Shuttle Radar Topography Mission for terrain elevation, Copernicus for land and vegetation characteristics, WorldPop for human population distribution, and PROBA-V for vegetation monitoring at moderate resolution.

The scale of the resulting feature dataset is worth appreciating. By sampling 13,000 spatial points across the campus and characterizing each with 40 features, the pipeline produces a rich multivariate description of the landscape. The third stage adds distance vectors, computing how far each sampled point lies from every digitized natural and infrastructural element. Distance to roads, distance to buildings, distance to water, and similar measures are among the most consistently informative predictors in the wildlife-crime literature, because they operationalize the trade-offs that poachers implicitly make between accessibility and concealment. Encoding these distances explicitly for every point means the dataset captures the accessibility structure of the entire campus, not just coarse regional averages.

The fourth and fifth stages are where community knowledge enters the pipeline, and they represent the most novel methodological contribution. Rather than asking respondents to classify arbitrary parcels, the researchers first applied K-Means clustering to group the landscape into spatially coherent regions, then asked a small panel of eleven to twelve community respondents to label each cluster. Clustering before elicitation reduces the cognitive burden on participants, a technique that aligns with prior work on incorporating ranger knowledge into spatial models, where the number of units a human must evaluate directly constrains the quality of the answers. The ternary labeling scheme is equally important: respondents could mark a region as poaching, non-poaching, or unknown. This explicit third category preserves epistemic honesty, preventing respondents from being forced into confident judgments about areas they genuinely know little about, which would otherwise inject noise disguised as data.

Label propagation, the fifth stage, then addresses the inherent sparsity and uncertainty of the community labels. Because only a small number of people labeled a limited set of clusters, the raw annotations cover only a fraction of the 13,000-point dataset. Propagation techniques spread the labeled information to unlabeled points based on spatial and feature similarity, while the unknown category allows uncertainty to be carried forward rather than collapsed into a false binary. The output is a continuous hotspot map in which different parts of the campus carry different levels of poaching risk confidence, and this map becomes the spatial prior that sensor-placement optimization was missing.

The validation strategy deserves attention because it tackles the obvious objection to community-sourced data: that non-expert recollections might be arbitrary. The descriptive analysis found that areas labeled as poaching hotspots exhibit ecologically coherent patterns, corresponding to locations with reduced human presence, denser vegetation cover, greater distance from buildings and infrastructure, and adjacency to transport corridors. These characteristics mirror findings from decades of wildlife-crime research in formally monitored protected areas, suggesting that the community respondents, whether consciously or not, encoded the same landscape logic that professional patrol data reveals elsewhere. In other words, the collective memory of residents appears to function as a compressed, informal record of the same underlying spatial processes that generate poaching risk.

The machine-learning confirmation strengthens this interpretation further. When classifiers were trained on the community-derived labels using the 40 satellite and distance features, the labels proved spatially learnable, with a K-nearest-neighbors model achieving a weighted F1 score of 0.96. A score this high indicates that the boundary between poaching and non-poaching areas, as drawn by community memory, aligns closely with sharp gradients in the environmental feature space. If the labels were random or idiosyncratic, no classifier could achieve such performance, because there would be no consistent relationship between the labels and the measurable characteristics of the landscape. The high F1 score therefore functions as an indirect quality check on the elicitation process itself, demonstrating that community knowledge, once structured, behaves like a usable ground-truth surrogate.

The choice of K-nearest-neighbors as the top performer is also informative. KNN is a locally oriented, non-parametric method that classifies points based on the labels of their closest neighbors in feature space. Its strong performance suggests that poaching and non-poaching regions form compact, well-separated neighborhoods in the 40-dimensional feature space, consistent with the idea that risk is driven by local environmental context rather than broad regional trends. For practitioners, this also implies that relatively simple, interpretable models may suffice for cold-start hotspot mapping, reducing the computational and expertise barriers for resource-limited conservation teams.

The downstream application of the hotspot map connects this work to the broader architecture of modern conservation technology. The derived risk surface guided the design and deployment of a fog-based, energy-efficient IoT monitoring system, described in detail in a companion paper. Fog computing places processing capability close to the sensors themselves, allowing devices to filter, aggregate, and react to data locally rather than transmitting everything to a distant cloud. In a large natural area with unreliable network coverage and constrained power, this architectural choice conserves energy and bandwidth, both of which are scarce resources for battery-powered field devices. The hotspot map makes such a system viable by ensuring that the limited fleet of energy-efficient sensors is concentrated where the probability of detecting poaching activity is highest, converting an intractable blanket-coverage problem into a targeted deployment problem.

The generality of the pipeline is perhaps its most significant practical attribute. Every component relies on freely available resources: OpenStreetMap and QGIS for digitization, open satellite missions for feature extraction, and a modest community panel for labeling. Nothing in the workflow requires historical patrol archives, proprietary enforcement data, or expensive commercial imagery. This means the approach can be replicated at any conservation area where official records are absent but community knowledge exists, a condition that describes a large share of the world’s ecologically significant but formally unmonitored landscapes. Small reserves, institutional campuses, sacred groves, and community forests, which rarely attract the patrol-record infrastructure of major national parks, are precisely the settings where this cold-start pathway offers the most value.

There are also broader implications for how conservation science thinks about data. The study demonstrates that ground truth need not be institutional to be useful; it can be distributed, oral, and informal, provided it is elicited with appropriate structure. The ternary scheme’s explicit representation of uncertainty is a model for other knowledge-elicitation efforts, since it acknowledges the epistemic limits of human memory rather than laundering them through forced binary choices. As IoT and edge-computing technologies continue to mature, the binding constraint on wildlife monitoring is increasingly not the sensing hardware itself but the intelligence about where to sense. Work like this pipeline addresses that constraint at its source, showing that with careful geospatial structuring, the collective memory of a community can be transformed into a rigorous, statistically validated foundation for protecting the landscapes those communities know best.

Subject of Research: …

Article Title: Community-informed identification and analysis of wildlife poaching hotspots for targeted IoT deployment

Article References: Siyanwal, R., & Srirama, S. N. (2026). Community-informed identification and analysis of wildlife poaching hotspots for targeted IoT deployment. Frontiers in The Internet of Things, 5, Article 1893789. https://doi.org/10.3389/friot.2026.1893789

Image Credits: AI Generated

DOI: 10.3389/friot.2026.1893789

Keywords: Community-informed, identification, analysis, wildlife, poaching, hotspots, targeted, deployment, scientific research