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

1 Introduction

Targeted monitoring of large natural areas remains an unresolved challenge in wildlife conservation because it must be conducted over large, heterogeneous areas that typically have limited surveillance. The challenge of efficiently deploying IoT or wireless sensor networks (WSNs) over large regions is known to be NP-hard in the general case (Elhabyan et al., 2019). Metaheuristic approaches, such as genetic algorithms, particle swarm optimization, and ant-colony optimization approximate but not guarantee optimal sensor placement and rely on prior assumptions about target areas, monitoring priorities, or the spatial distribution of event likelihoods. While IoT/fog-based sensor networks (Srirama, 2024) can, in principle, detect wildlife poaching in near real time using acoustic sensing, edge computing, unmanned aerial systems, and real-time computer vision, their deployment at the required scale is still restricted by cost, energy availability, network limitations, and insufficient infrastructure. As a result, providing uniform sensor coverage over an entire landscape is both economically and operationally impractical. The key question is therefore not whether to deploy sensors but where they should be placed. In wildlife conservation contexts, such structured prior information is rarely available. Effective deployment depends on advance spatial insight into where illegal activities are most likely, yet many conservation landscapes lack the necessary historical patrol records, incident logs, ranger reports, or long-term monitoring datasets to generate this intelligence.

This leads to a cold-start conservation problem, in which monitoring systems must be designed and implemented without historical ground-truth data. The issue is particularly acute in developing regions and resource-limited contexts, but it is not confined to them. Ecologically significant urban forests, fragmented habitats, university or institutional campuses, and community-managed conservation areas also frequently face wildlife crime. In these environments, knowledge about poaching risk exists in fragmented and undocumented form through the collective memory of local communities, security staff, field researchers, and long-term residents.

The University of Hyderabad (UoH) campus, which serves as the empirical study site for this paper, is representative of this situation. The campus spans over 2,000 acres of ecologically diverse terrain in Telangana, India, and actively supports populations of mammal, bird, and reptile populations. Campus staff, security personnel, and researchers concur that poaching has taken place across several areas of the campus over many years. The sole available ground truth, however, consists of informal, collectively held knowledge accumulated and undocumented, from people who reside, work, or conduct field studies in this landscape.

Transforming such knowledge into a structured, machine-readable, and spatially explicit dataset is the primary technical challenge addressed in this paper. Existing wildlife-crime prediction systems, such as PAWS, SMART-based analytics, and comparable machine-learning frameworks generally assume the availability of ranger patrol logs, historical encounter data, or systematically curated monitoring records. These systems have demonstrated that poaching is not randomly distributed in space and can be modeled using environmental and anthropogenic predictors. However, they do not fully address cold-start environments.

Recognizing the value of biodiversity and the need to move beyond reactive security measures, the UoH supported the development of a systematic, data-driven approach presented here. The objective of this work was to develop a complete geospatial pipeline to generate wildlife poaching hotspot maps directly from community-sourced spatial knowledge. In this paper we demonstrate such pipeline that integrates OpenStreetMap-based spatial digitization, multi-source satellite feature extraction, infrastructure-distance modeling, community-driven spatial labeling, uncertainty-aware label propagation, and hotspot construction within a single operational workflow. In contrast to purely binary classification methods, the elicitation method in this pipeline explicitly encodes uncertainty through ternary labeling scheme consisting of “poaching”, “non-poaching”, and “unknown” categories. This avoids forcing the respondents to classify regions about which they possess insufficient knowledge. This work provides the spatial foundation that has been missing so far for target conservation actions.

Our study further shows that community-derived spatial labels are neither arbitrary nor random; instead, they yield hotspot patterns that are both ecologically coherent and statistically differentiable. Descriptive analyses indicate that regions characterized as vulnerable to poaching systematically correspond to areas with reduced human presence, denser vegetation cover, greater distance from buildings and infrastructure, and adjacency to transportation corridors. These attributes closely mirror established findings in the wildlife-crime literature, which document that illicit activities in protected areas tend to exhibit consistent and predictable relationships with local environmental and landscape features.

Existing anti-poaching research has already demonstrated the value of patrol records, ranger knowledge, environmental co-variates, spatial risk models, and operational decision-support systems. In particular, prior studies have used expert knowledge in spatial models and have used K-means defined regions to reduce the burden of ranger elicitation. The unresolved problem addressed here is therefore narrower, which is, small conservation sites may need an initial spatial-prioritization layer even when they do not yet possess a structured patrol or incident database.

Beyond the hotspot mapping itself, the derived intelligence motivated the design of a fog-based, energy-efficient wildlife monitoring system, presented in detail in the companion paper (Siyanwal et al., 2025) and briefly outlined in Section 5 of this paper. The main contributions of this work are:

• A fully cold-start geospatial pipeline for generating wildlife poaching hotspot maps in settings where no historical incident data are available.
• A procedure to spatially digitize a land parcel (in our case, the UoH campus) from scratch for research purposes using OpenStreetMap ¹ and QGIS ² .
• A multi-source satellite-based feature extraction pipeline that integrates SRTM, Copernicus, WorldPop, and PROBA-V data, yielding ∼13,000 points with 40 features.
• A community knowledge elicitation framework and a ternary (1/0/-1) labeling strategy that explicitly represents epistemic uncertainty.
• An extensive descriptive analysis demonstrating that community-identified hotspot regions possess ecologically consistent and statistically distinguishable environmental characteristics.

The remainder of this paper is structured as follows. Section 2 surveys the relevant literature and clearly situates this study in relation to existing work. Section 3 describes the cold-start geospatial data pipeline. Section 4 presents the descriptive and machine-learning-based validation of the resulting hotspot dataset. Section 5 introduces the fog-based IoT monitoring architecture enabled by the hotspot map. Section 6 discusses the empirical findings, and Section 7 concludes the paper.

2 Related work

2.1 The large-scale IoT deployment problem

The problem of deploying IoT or Wireless Sensor Network (WSN) systems efficiently in a large area, in the general case, is established as NP-hard (Elhabyan et al., 2019). Metaheuristic approaches, such as, particle-swarm optimization, and ant-colony optimization provide near-optimal solutions to the placement problem in WSNs but remain susceptible to local optima, slow convergence, and stagnation (Sun et al., 2020). Parmar et al. (2024) showed that sparse IoT deployment failed to capture localized pollution hotspots, highlighting the importance of effective sensor placement. Empirical studies across multiple domains demonstrate that placement strategies influence sensing quality, network topology, communication delay, and infrastructure cost. For example, in urban air-quality monitoring system, Boubrima et al. (2017) demonstrated that optimized deployment can significantly reduce the number of required sensors while maintaining monitoring objectives; in smart-city WSN application, Alablani and Alenazi (2020) reports ∼30% improvement in both sensing coverage and end-to-end delay using triangulation-based deployment strategy. Collectively, these studies indicate that deployment planning, that is, the question of where to deploy is a foundational factor in the effectiveness of large-scale IoT or WSN systems.

2.2 Wildlife poaching as a quantified, large-area monitoring problem

2.2.1 Scale, severity, and spatial structure

The UNEP-INTERPOL 2016 report ³ states that environmental crime is the fourth largest illicit trade estimated at USD 91–258 billion annually. Among these crimes, illegal wildlife trade is estimated around USD 7–23 billion annually (7%–9%). Chase et al. (2016) report that a renewed poaching outbreak beginning around 2005 has led to the death of an estimated 30,000 elephants per year and documented a decline of 144,000 Savannah elephants between 2007 and 2014, shrinking their population roughly at 8% per annum. Ferreira et al. (2015) in 2016, projected detectable rhinoceros population decline in Kruger National Park under continued poaching pressure, and Hauenstein et al. (2019) demonstrated that poaching intensity correlates with local poverty, national corruption, and ivory price dynamics. Ghoddousi et al. (2022) used logbook data to map poaching pressure across a large protected area and found it concentrated where prey availability is highest.

Although most research on poaching severity focuses on large protected areas in Africa, the same underlying conditions, such as, local poverty, easy access via unmonitored borders, and spatially concentrated risks, also characterize peri-urban and campus conservation areas. In South Asia, the poaching of birds, reptiles, and small mammals for subsistence, the pet trade, and traditional medicine is widely acknowledged as a problem. The UoH campus is a large-scale example of this situation. It is ecologically rich, experiences active poaching, and largely lacks formal patrolling infrastructure. The same structural monitoring gaps documented in large reserves are therefore present on the UoH campus as well.

2.2.2 Community-based conservation monitoring

Community-based monitoring provides a complementary alternative to conventional conservation surveillance. In 2009, Danielsen et al. (2009) introduced a foundational typology of natural resource monitoring systems and argued that locally driven monitoring (involving local communities) can deliver sustainable, context-appropriate approaches in resource-constrained environments. Critchlow et al. (2015) analyzed ranger-collected data using spatially explicit occupancy models to examine spatial and temporal patterns of illegal activities in protected areas. They observed substantial spatio-temporal variation in the occurrence of illegal activities and showed that past locations of such activities were among the strongest predictors of where they would occur in subsequent years. Thiault et al. (2020) demonstrated spatial poaching risk modeling by applying Boosted Regression Tree (BRT) to patrol data and incident reports, although not community-based, this work is analogous to our work with distinction that formal records do not exist in the UoH setting.

2.3 Existing technical solutions

2.3.1 Machine learning for poaching prediction

The Protection Assistant for Wildlife Security (PAWS) (Xu et al., 2020) uses Gaussian processes and historical patrolling data to generate uncertainty-aware poaching risk predictions and prescribes patrol routes, their field evaluations reported 30% increase in snare detection on average. Guo (2021) extended PAWS using satellite-derived remote sensing features to improve prediction capability in under-resource parks, the framework relied on historical illegal activity labels for supervised learning. Earlier systems, such as CAPTURE (Nguyen et al., 2016) relied on Bayesian-network-based framework. This paper takes the core architectural insights from Gurumurthy et al. (2018) who demonstrated the K-means Lloyd (1982) clustering to spatial features in the Huang Ni He National Reserve, China. However, our work is applied in qualitatively harder setting: no professional rangers, no prior records, non-expert community respondents whose limits must be handled explicitly, and a downstream application of a physical IoT sensor placement rather than algorithmic patrolling route. Collectively, these approaches assume the availability of prior labeled patrol observations and do not fully address cold-start environments in which no historical incident records exist.

2.3.2 Sensor networks, acoustic monitoring, and drones

Passive acoustic monitoring has been investigated as a tool for anti-poaching surveillance in forested environments. Hill et al. (2018) introduced the AudioMoth platform for gunshot detection based on acoustic signals. Their results showed that gunshots in tropical rain-forests could be detected at distances of up to 500 m, and up to 1 km under continuous monitoring conditions. Mulero-Pázmány et al. (2014) assessed the use of remotely piloted aircraft systems for rhinoceros anti-poaching operations and highlighted practical operational constraints that limit their field deployment. PoachNet (Hamed et al., 2024) integrates deep learning with Semantic Web reasoning to predict poaching likelihood from GPS trajectories of tagged elephants. Bondi et al. (2018) proposed SPOT, a UAV-mounted thermal infrared detection system that enables near real-time automated identification of poachers and wildlife using Faster R-CNN–based computer vision models. Camera-trap research, such as Glover-Kapfer et al. (2019) and Burton et al. (2015), has underscored that imperfect detection, camera positioning, sensor properties, and sampling design can profoundly influence wildlife monitoring inferences and the reliability of surveys. The Fog-IoT architecture described in the companion paper (Siyanwal et al., 2025) tackles energy and communication limitations but presumes that deployment zones are already known, the knowledge this paper seeks to substantiate.

The Spatial Monitoring and Reporting Tool (SMART) ⁴ is widely used extensively to collate, manage, and analyse ranger patrol data collected in protected areas. Dobson et al. (2019) observe that patrol performance is generally assessed using historical patrol records and encounter information. However, these systems rely on accumulated patrol observations, which restricts their usefulness in contexts where no previous monitoring data are available.

2.4 Simple, ternary community elicitation

The elicitation instrument consisted of a single, straightforward question with three response options. It was based on Satisficing theory (Krosnick, 1991), which argues that when questions place high processing demands on respondents, they tend to switch to cognitively simpler response strategies. Experimental evidence from Lenzner et al. (2010) demonstrates that linguistically complex survey questions increase respondents’ cognitive load and are linked to markers of satisficing behavior. In addition, Dolnicar et al. (2011) showed that simple binary response formats are quicker to answer, require less cognitive effort, and provide reliability on par with multi-category Likert scales. Taken together, these studies support the use of low-burden elicitation formats.

Including an “unknown” or “I have no information” category (−1) alongside 1 and 0 directly addresses the failure mode that arises when respondents are compelled to make judgments that exceed their actual knowledge. Torre et al. (2019), in a citizen-science classification study, showed that providing an “I do not know” option can improve accuracy relative to setups that allow only binary choices. Similarly, Cao et al. (2021) argued that standard semi-supervised learning methods are inherently based on a closed-world assumption, namely, that unlabeled test samples belong only to classes observed during training, and that the presence of previously unseen classes diminishes the effectiveness of these methods. In the setting of this paper, a grid cell without a recorded poaching event may indicate true absence of poaching or merely a lack of observation. Interpreting all such cells as confirmed non-poaching injects uncertainty into the negative class labels. The −1 category makes this uncertainty explicit.

Our work differs from the most close related approaches in three aspects. First, it is designed for a data-scarce, cold-start setting and does not depend on pre-existing structure database of patrol encounters or incident locations. In contrast, systems such as CAPTURE and PAWS learn from accumulated patrol records, while PoachNet relies on tagged-animal telemetry. Second, the study incorporates spatial knowledge from a heterogeneous institutional community rather than relying exclusively on professional ranger assessments, enabling respondents to identify recognizable areas without providing precise indecent coordinates or detailed even history. Third, it retains “unknown” coded as − 1 as a separate knowledge-coverage response rather than forcing respondents into binary choices. These distinctions position the proposed pipeline as a practical cold-start workflow for identifying candidate monitoring areas where structures incident data are unavailable. Table 1 summarizes the key difference between the present study and the relevant prior approaches.

TABLE 1. Comparison of the study with the prior work relevant to poaching evidence, community and ranger monitoring, patrol allocation, predictive anti-poaching systems, and elicitation uncertainty.

3 Methodology

3.1 Study area

The University of Hyderabad (UoH), established in 1974, is a central university located in Gachibowli, Hyderabad, Telangana, India. The campus covers approximately 2,000 acres and is recognized as an eco-sensitive zone. It supports a rich diversity of over 455 species of flora and fauna. The university has at least 10 recorded populations of mammals such as Spotted Deer, Wild Boars, Hares and Rabbits, Indian Grey Mangoose, Porcupines, and Bats. The university is home to over 220 species of resident and migratory birds and shelters approximately 15 species of reptiles. The terrain includes four major water bodies, three wetland zones, dry scrubland, grassland, forest patches, and approximately 200 built structures. Despite confirmed and community-acknowledged poaching incidents spanning multiple decades ⁵ , no extensive incident public-record exists, such as patrol logs or wildlife survey with incident mapping.

3.2 Pipeline overview

The pipeline transforms a blank spatial record into a labeled hotspot dataset in five sequential stages (Figure 1). Each stage is described in detail below:

3.2.1 Stage 1: Spatial digitization

At the outset of this study in 2022, no digital spatial representation of the UoH campus was available in any public GIS repository (the spatial data used here were also acquired in 2022). Consequently, the campus had to be mapped from the ground up using OpenStreetMap (OSM), QGIS, and ArcGIS. All spatial layers were stored in the EPSG:4326 coordinate reference system (WGS 84) to maintain compatibility with all satellite datasets employed in the later stages. Using the OSM web editor, all identifiable campus feature was annotated in accordance with the OSM map features schema: four lakes, three wetland areas, all buildings (including academic schools, hostels, administrative offices, the powerhouse, etc.), the complete road and path network (highway tags applied: residential, track, service, secondary, primary, footway, pedestrian), the overall campus boundary polygon, the high-tension transmission line corridor, and natural land-cover classes (grassland and forest patches). The tagged features were then imported into QGIS via the QuickOSM plugin, which enabled direct retrieval from the OSM Overpass API. To validate the spatial accuracy of the resulting digital map, we overlaid it on Bing Aerial imagery. This comparison confirmed that all mapped features aligned with their corresponding satellite structures, verifying the geometric correctness of the dataset. The individual digitized spatial layers are presented in Figures 2a–e, while the complete digitized campus map is presented in Figure 2f.

3.2.2 Multi-source satellite feature extraction

Once the campus was spatially digitized, the subsequent task was to assign a numerical profile to every location. Poaching risk is unevenly distributed across the landscape and is shaped by factors such as topography, vegetation, land cover, human population density, and levels of urban development. To represent these spatial patterns, a regular point grid with a spacing of 0.0004° (about 43 m) was created in QGIS. After limiting the grid to the study area and removing out-of-range points, 12,987 valid grid points remained, each corresponding to a 43 m × 43 m terrain cell. Four satellite datasets: Shuttle Radar Topography Mission (SRTM) Digital Elevation Model (DEM) (Farr et al., 2007), Copernicus Global Land Cover (Buchhorn et al. (2020)), WorldPop ⁶ , and PROBA-V Top of Canopy (2014–2020) ⁷ —were obtained through Google Earth Engine (GEE) ⁸ and the United States Geological Survey (USGS) EarthExplorer ⁹ platform, both of which provide free access for academic use. The grid spacing was chosen to be consistent with the coarsest input layers, including WorldPop and Copernicus (both at 100 m resolution), to avoid implying a higher spatial precision than that supported by the underlying data. The four satellite products sampled at each grid point are summarized in Table 2, together with the derived features they contribute. Raster values for all grid points were extracted using the Add Surface Information tool in ArcMAP. Following this processing step, the resulting dataset contained 12,987 records and 26 features.

TABLE 2. Satellite datasets used for feature extraction.

3.2.3 Stage 3: Distance vector computation

A poaching event requires both a suitable habitat and an accessible pathway that allows a poacher to enter, operate, and exit. The proximity of grid point to roads, campus boundaries, buildings, power infrastructure, and other built-up features captures information about accessibility that cannot be inferred from satellite reflectance data alone. Stage 3 therefore augments the 26 satellite-derived features with 14 additional distance-based features.

These distance-based features were generated with the NNJoin (Nearest Neighbor Join) plugin in QGIS. NNJoin ¹⁰ calculates the Euclidean distance from each grid point to its nearest corresponding feature. We applied NNJoin to eight types of infrastructure: the highway network (producing both distance and highway category), campus boundary, buildings, land-use zones (distance plus land-use category), leisure areas (distance plus leisure category), natural features (distance plus natural category), office buildings (distance plus office name), and power infrastructure (distance plus power category). In total, this procedure contributed 14 additional features. Figure 3 illustrates the spatial grid and the computation of distance vectors from a grid point to selected mapped features.

3.2.4 Stage 4: Community knowledge elicitation

Now, the dataset is rich in environmental and infrastructure features but contains no information about where poaching activities have actually occurred. This is the challenge of cold-start setting. Without historical records, the only available ground-truth information must come from residents who carry the relevant knowledge. The majority of the students in UoH are enrolled in 2-year Master’s programmes, a duration too short for most students to have witnessed poaching incidents in the campus. Therefore, the respondents identified were security team members, long-tenure faculty, and research scholars. The elicitation protocol was designed to extract informal, memory-based spatial knowledge from respondents who could not be expected to recall specific incidents with precision but who could plausibly recognize spatial zones and report whether poaching had occurred there. The survey was conducted across three community groups described in Table 3.

TABLE 3. Community respondent groups and their nature of knowledge contribution to the elicitation protocol.

Where respondents provide conflicting knowledge, the following decision rules were applied:

• A poaching label (1) from any respondent overrides a non-poaching label (0).
• A non-poaching label (0) overrides an unknown label (−1).
• A poaching label is never downgraded.

Asking respondents to reason about 12,987 individual grid points is cognitively infeasible. We need to present this information in a way that people can think about and move through the landscape, for this, we employed K-Means clustering that partitioned 12,987 points to spatially coherent clusters, each cluster is an area where the terrain, land cover, and infrastructure are internally consistent. The decision for the number of clusters was calculated by the Elbow method applied to Within-Cluster Sum of Squares (WCSS)Thorndike (1953) curve and discerned to k = 80 because it produced clusters of approximately 25 acres in size on an average. Each cluster was rendered as a distinct color on a printed map given to the respondents.

The question administered per cluster: “Given this area in the campus, do you know whether poaching has taken place here?“. The respondent then would mark the cluster with either of the three categories: “1” (poaching has occurred here), “0” (the respondent reported that poaching has not occurred), and “-1” (I have no information about this area). Instead of forcing the respondents to classify every cluster either poaching or non-poaching, we introduced third category of “unknown” as well. If there were only two categories, clusters for which the respondents have no knowledge would be labeled 0 b y default, hence polluting the non-poaching class with uncertain observation. The “-1” category reflects open-world assumption instead of forced binary labeling. With cluster label assigned, every grid point inherited the label of the cluster to which it belonged to. The resulting label distribution was as follows: 7,438 points labeled 0 (non-poaching, 57%), 1,691 labeled 1 (poaching, 13%), and 3,858 labeled −1 (unknown, 30%). The dataset construction pipeline is represented in Algorithm 1 and the dataset features along with their description is presented in Table 4.

TABLE 4. Dataset variable descriptions. The complete dataset comprises 12,987 spatial grid points across the University of Hyderabad campus, each described by Longitude and Latitude, grid point position (inside or outside UoH), 40 features, and a ternary poaching label.

Algorithm 1 Poaching prediction dataset construction pipeline.

3.2.5 Stage 5: Hotspot map generation

Till stage 4, the dataset consisted of spatial grid points grouped into 80 clusters according to similarity in their environmental and infrastructural features. During community elicitation, a response assigned to a cluster was propagated to all grid points belonging to that cluster. Consequently, when a cluster was labeled 1, geographically separate locations sharing the same cluster membership and feature profile also inherited the poaching label, even if those locations had not been individually identified by the respondents. This feature-based propagation enabled the identification of additional candidate poaching-prone areas with characteristics similar to the community-identified locations. The label 1 grid points were subsequently visualized using QGIS heatmap renderer, as shown in Figure 4 (right). In this study, a hotspot therefore represent a spatial concentration of points belonging to clusters assigned a poaching label. The resulting map is a community-informed, feature-propagated prioritization map rather than a calibrated estimate of poaching probability.

K-means was used as a practical region-construction step to reduce the number of map units presented during community elicitation. It was not used to infer poaching occurrence. The resulting 80-region partition was treated as the fixed elicitation instrument used in this case study. Figure 4 uses QGIS heatmap renderer. A hotspot in this figure mean the region was assigned a community response of 1. Therefore, based on the collected and derived features, it also shows hotspots that are poaching prone but were not assigned a label 1 b y the community.

3.2.6 Data preprocessing

We analyzed the data and found no duplicates. Missing values were observed in the Nearest Office attribute, accounting to 0.04% of the total observations, hence negligible. To improve readability and consistency in the dataset, the original label variables: 1 and 0 were converted to Poaching Area and Non-Poaching Area, respectively. Apart from these, the following transformations were done on some features:

• Rows labeled as “-1” were dropped (only for descriptive analysis).
• Derived a feature from Copernicus Land Cover called Cover Type.
• Cover Type Groups: In Cover Type, various categories were clubbed together, such as “Shrubs”, “Herbaceous Vegetation”, “Vegetation Land”, “Evergreen Broad Leaf Shrubs and Grasslands”, “Seasonal Needle Leaf Shrubs and Grasslands”, “Land with water and small vegetation”, and “Moss” were grouped as “Vegetation”; “Evergreen Needle leaf canopy”, “Evergreen Broadleaf tree canopy”, “Seasonal needle leaf canopy”, and “Seasonal Broad leaf canopy” were grouped as “Tree Canopy”; “Closed forest” and “Unknown Closed forest” were clubbed under “Closed Forest”; “Open forest” and “Unknown Open forest” were clubbed under “Open Forest”; “Permanent Water Bodies” and “Oceans/Seas” were grouped as “Water Bodies”.
• Population_Count: Derived from WorldPop feature by creating the following bins: “No Population”, “1–20”, “21–40”, “41–60”, “61–80”, “81–100”, and “100+“.
• ‘Highway_Dist (meters): Derived from Highway Distance feature by creating the following bins: “Less Than 1 m”, “1–30”, “31–60”, “61–90”, “91–120”, and “120+“.
• Nearest HCU Boundary Distance, Nearest Leisure Distance, Nearest Natural Distance, and Nearest Office Distance were transformed into the bins: “Less Than 1 m”, “1–30”, “31–60”, “61–90”, “91–100”, “101–500”, “501–1000”, and “1000+“.
• Nearest land-use Distance (meters) was transformed into the following bins: “Less Than 1 m”, “1–30”, “31–60”, “61–90”, “91–100”, “101–500”, “501–1000”, and “501–1000”, “1001–1050”, “1050–2000”, and “2000+”.

4 Results

The results summarize patterns within the community-label spatial dataset. Percentages and mean feature values are reported descriptively to show how mapped environmental and infrastructural characteristics vary across the response categories. The primary analytical tool used in this study was Bivariate analysis. The metric reported throughout is the percentage of observations in each feature category. Table 5 summarizes the descriptive statistics of all continuous environmental features across the three label categories, providing the foundation for the Bivariate analysis discussed in the following subsections.

TABLE 5. Descriptive statistics of environmental variables across poaching labels.

4.1 Spatial influence of the university boundary

The obvious spatial asymmetry was identified with respect to the university’s administrative boundary, where, the observations were distributed equally between the interior (47%) and the exterior (53%) of the university but the spatial distribution shows that approximately 70% of poaching observation were located inside the university with only 30% situated outside. This disparity is further quantified by conditional probability analysis in which 27% interior locations carry a poaching label compared to 11% outside.

Inside-campus locations are 2.45x more likely to be identified as poaching areas. The campus interior offers lower urbanization pressure, lower surveillance density, hence, establishing wildlife corridors that poachers may have learned over time. The external regions that have experienced greater urbanization pressure, higher infrastructure density, and elevated human disturbance have collectively reduced both wildlife presence and poaching opportunity. The interior-exterior asymmetry is consistent with an observation made by Critchlow et al. (2015).

4.2 Topographical characteristics

The elevation analysis in Figure 5 indicates that poaching sites occur within a slightly higher and more variable elevation range, with a median near 592 m and an interquartile range of 578–630 m above sea level. In contrast, non-poaching sites are more tightly clustered around a slightly lower median of about 587 m, with most values concentrated in a narrower 581–596 m band, although a few outliers reach up to 652 m. This greater elevation spread at poaching locations suggests that incidents are not restricted to a single topographic band, but are more likely to occur across mid-elevation slopes where terrain remains passable, canopy cover is moderate, and wildlife movement corridors are relatively open. The slope analysis supports this reading: poaching and non-poaching sites exhibit nearly the same slope distributions, sharing a median of roughly 2 ° and an interquartile range of 2 − 4 °, which implies that slope gradient alone does not differentiate between them. Instead, poaching sites show a lower upper limit in extreme slope values, rarely exceeding 10 °, whereas non-poaching locations sometimes reach gradients beyond 35 °, suggesting that human intruders generally avoid the steepest terrain regardless of wildlife occurrence.

4.3 Population density

The median WorldPop (Figure 6) count for areas labeled as poaching sites is 0, and roughly 75% of all recorded poaching incidents occur in cells with no population. A bivariate analysis further indicates that 23% of zero-population cells are classified as Poaching. Taken together, these results indicate a strong link between poaching activity and locations with minimal human presence, aligning with the expectation that low-surveillance settings enable covert behavior. Within the campus, zero-population zones therefore constitute the top-priority surveillance locations for IoT deployment.

4.4 Transportation infrastructure and accessibility

Poaching areas, as shown in Figure 7 (Left), are typically situated farther from highways than non-poaching regions, with median distances of about 36 m and 26 m, respectively. The upper quartile for poaching locations extends to roughly 91 m from transport corridors, compared with about 64 m for non-poaching locations, indicating that poaching tends to occur in areas that are more spatially isolated. Heatmap analysis also reveals that the strongest association with poaching is found in zones located more than 120 m from highways. Collectively, these results show a monotonically increasing relationship ( 0.14 → 0.16 → 0.18 → 0.19 → 0.22 → 0.28 ) indicating that the further from highways, the higher the poaching probability.

Highway-type analysis in Figure 7(Right), further shows that poaching activity is disproportionately associated with primary roads and track-type routes. Within university, track routes consist of unpaved internal roads that traverse ecologically sensitive regions while remaining relatively less monitored than other urban infrastructure. Residential and service roads show moderate association, whereas pedestrian-oriented and tertiary road categories exhibit little or no association.

4.5 Human infrastructure and surveillance proximity

As shown in Figure 8, we found that the poaching observations were concentrated in regions located at greater distance from buildings, whereas, the areas characterized by sparse built environments showed the highest proportion of poaching labels. Median building distance for poaching locations is 145 m vs. 61 m for non-poaching (more than double). Locations > 100 m from any building represents one of the sensor placement priority.

A similar trend was observed with respect to office proximity. The median distance of a poaching-labeled area from an office is 1,199 m vs. 814 m for non-poaching areas. Locations situated more than 1,000 m from the nearest office showed the highest proportion of poaching labels. The office represent the highest activity density structure, their spatial influence extends further than other buildings.

A counter-intuitive pattern was found with power-line proximity. As shown in Figure 9, Poaching labeled areas are closer to power infrastructure (median 175 m vs., 418 m for non-poaching). The high-tension power line corridors in the campus create a cleared linear zones through the campus interior, functioning as an internal route. It allows movement without road use and provides staging ground.

Poaching locations are situated at approximately double the distance from the nearest building compared to non-poaching. This structural separation strengthens the hypothesis that poaching tends to occur in areas with limited administrative or institutional oversight.

4.6 Land cover

As shown in Figure 10, Evergreen Broadleaf Tree Canopy areas carry 100% poaching label rate. Every location in their land cover class was identified as a poaching area. Closed forest and Open forest also showed high proportions. Seasonal Broadleaf Shrubs and Grassland areas have non-zero poaching labels.

4.7 Land-use

In Figure 11, Land-use analysis indicated that proximity to farmland is the strongest land-use correlation to poaching activity among all categories examined. This association may reflect increased wildlife movement near agricultural boundaries driven by food availability, which in turn increases human-wildlife engagement and creates opportunity for poaching and possibly illegal harvest.

Leisure-site proximity also exhibited a notable relationship with poaching risk. Locations falling within 101–500 m of recreational facilities displayed higher poaching probability with sports centers in the university (in the middle on the poaching hotspot map) showing the strongest association. This pattern may be attributed to the increase in human-wildlife interaction introduced by recreational infrastructure at habitat boundaries.

4.8 Summary of descriptive findings

The descriptive analysis identifies a multi-factor spatial signature associated with poaching-prone areas. The proximity to primary transportation corridors, isolation from buildings and administrative infrastructure, association with farmland, and accessible mid-elevation terrain are directly proportional to poaching activities. In essence, poaching activities are influenced by ecological accessibility and reduced human oversight rather than by any single environmental driver. A composite spatial profile of UoH poaching hotspot is represented in Table 6.

TABLE 6. Characteristic feature values observed in poaching-prone areas within the UoH campus.

The identified spatial patterns provide an empirical foundation to informed IoT system placement, patrol-route planning, and prioritization of surveillance resources in high-risk zones within wildlife monitoring systems.

4.9 Validation using machine learning

The descriptive analysis above establish that the community labels are ecologically coherent. A second validation is provided by machine learning classification. We wanted to check if spatial features in the dataset captured genuine signals about poaching status then the classifiers trained on the dataset should achieve meaningful better-than-chance performance. Four classifiers: Decision Tree, K-Nearest Neighbor (KNN), Logistic Regression, and Naive Bayes were used to evaluate the dataset. All three label classes (1, 0, −1) were included in classification. The dataset was partitioned into 70% training and 30% test sets with fixed random seed of 42 for reproducibility. Evaluation metrics used were precision, recall, and F1 score per class, with weighted average F1 as the primary metric (given the class imbalance). The results of the classifiers on all 40 features are shown in Table 7.

TABLE 7. Performance comparison of classification algorithms across different labels.

To assess whether a reduced feature set (feature ablation) could maintain classification performance, we applied one-way ANOVA using F-statistics as the selection criterion (features with well-separated class means receive higher F-scores). The selected feature subset was then used to retrain and re-evaluate all classifiers. Table 8 presents the results. The top 10 features are: Nearest Office, BLUE_PROBA_V_2020, Power Distance, Nearest Boundary Distance, WorldPop, Nearest Building Distance, Nearest Leisure Distance, NIR_PROBA_V_2020, BLUE_PROBA_V_2019, and Nearest Office Distance.

TABLE 8. Performance comparison of classifiers after feature ablation and re-evaluation.

Feature selection via ANOVA marginally reduces performance for Decision Tree (−0.02) and KNN (−0.02), and more substantially for Logistic Regression (−0.09) while Naive Bayes shows a marginal decrease (−0.01). The dominant finding is that KNN on all features achieves the highest performance at weighted F1 = 0.96, with Decision Tree close behind at 0.95.

4.10 Independent institutional corroboration

After producing the hotspot map, the pipeline’s spatial predictions were informally shared with senior institutional leaders who were not part of the respondent group. They agreed that the pipeline’s identified hotspot areas aligned with their own sense of where poaching risks are concentrated on campus. This amounted to confirmation through conversation rather than a systematic review. As such, the support provided by leadership is best characterized as convergent community insight from a second, independent source, rather than as fully external validation.

5 Fog-based IoT framework enabled by the hotspot map

The poaching hotspot zones identified through community elicitation and kernel density estimationSilverman (1986) provides a direct basis for deciding where to install the IoT system. This section outlines the fog-based, energy-efficient IoT framework developed at UoH. A complete description of the framework is provided in the companion paper: Siyanwal et al. (2025). The following subsections detail the system architecture, its energy and latency characteristics, and the simulation-driven device selection strategy that enabled deployment at this scale.

5.1 Framework architecture

The framework adopts a three-tier architecture, as depicted in Figure 12. The Sensing Layer is composed of Motion Sensing Units (MSUs), each integrating four Passive Infrared (PIR) sensors oriented in opposite directions and interfaced with an ESP8266 NodeMCU microcontroller. Each MSU consumes approximately 0.36 W, making it suitable for solar-powered deployment in remote areas without access to the power grid. The Fog Layer contains the Actuating and Processing Units (APUs), where each APU includes a night-vision camera mounted on a servo motor and a single-board computing platform (Raspberry Pi 4 B). When an MSU detects motion, it transmits its unique identifier (UID) to its corresponding APU manager using the ESP-NOW peer-to-peer communication protocol. After receiving the UID, the APU orients the camera toward the MSU’s direction and captures a 1080p image. The APU’s compute node then processes this image and forwards it to the Hub Layer over a SoftAP WiFi connection using Apache Kafka. A single servo motor can pivot the camera through an angle of up to 180 °, removing the necessity for multiple fixed cameras. The Hub Layer hosts the Apache Kafka broker, subscribes to the APUs, and asynchronously processes images coming from the different APUs.

5.2 Device selection through simulation

Designing an IoT framework involves hardware trade-offs among latency, energy use, and cost, which are challenging to assess purely analytically. To overcome this, we evaluated 2,800 simulated architectural configurations using the iFogSim2 (Mahmud et al., 2022) simulation tool, systematically varying the computing devices at both the APU and hub levels across ten candidate platforms: Raspberry Pi 2, 3, 3 B+, 3 A+, 4 B (4GB and 8 GB), NVIDIA Jetson Xavier NX, Intel NUC 11, LattePanda Delta, and ASUS PN50. For the UoH deployment, the Raspberry Pi 4 B (8 GB) was chosen as it provided the basis of its balance across all three criteria. The resulting system delivered an end-to-end latency of 5 s in both dense and semi-forested environments. Figure 13 shows sample images captured by an APU.

6 Discussion

The principal finding is that spatial labels provided by community members yield a dataset whose poaching hotspots indicate a trade-off pattern between accessibility, concealment, and surveillance. Poaching-labeled locations were substantially more isolated from built infrastructure. For instance, the median distance from the nearest building was 145 m as compared to 61 m for non-poaching ones, while the corresponding median distances from offices were 1,199 m and 814 m, respectively. Although the median distance from the nearest road was only moderately greater for poaching locations (36 m vs. 26 m), the probability of a poaching label increased monotonically with the road distance (from 0.14 in the nearest distance interval to 0.28 beyond 120 m. These results align closely with, and extend, the established spatial analysis literature on wildlife crime. Plumptre et al. (2014) showed that illegal activities in protected areas follow non-random spatial patterns that can be modeled using environmental and human-related predictors, while Critchlow et al. (2015) demonstrated that different categories of illegal behavior are associated with distinct ecological and spatial niches within protected landscapes.

At the same time, road type produced an important contrast where the unpaved internal tracks showed the strongest association with poaching. This pattern is consistent with evidence that infrastructure can either facilitate illegal activities or increase the probability of detection, depending on its function and intensity of legitimate human use. For example, Moore et al. (2018) reported greater occurrence of poaching-related threats near roads and tourist trails but lower occurrence near ranger posts. Moreover, Thiault et al. (2020) showed that accessibility should not be interpreted in isolation and demonstrated that poaching risk is context-dependent. The UoH findings support these observations where, for wildlife poaching, the access is necessary but the most suitable ones are the concealed ones. This observation is strengthened by poaching locations being closer to UoH’s internal high-tension power line infrastructure (median 175 m vs. 418 m for non-poaching areas), which forms clear passage through the campus interior and is still concealed for general purpose use. The environmental characteristics of the identified areas further support the conclusion that the community-derived labels contain meaningful spatial structures. For instance, approximately 75% of poaching observations occurred in zero-population cells while 23% of all zero-population cells received a poaching label. Farmland-adjacent locations exhibited the strongest land-use association with 32% poaching-label rate. Locations 101–500 m from recreational infrastructure also showed elevated association with poaching.

The proposed pipeline is supported through three complementary forms of evidence evaluating whether the elicited labels are spatially meaningful, learnable, and operationally useful. The first line, as we have discussed, is ecological consistency confirmed with descriptive data analysis. The second line is machine learning predictability in which KNN achieved the highest F1 of 0.96, with Decision Tree at 0.95 on the three-class labeling problem. The performance of these classifiers further suggests that poaching risk emerges from interacting environmental conditions rather than simple linear associations. The third line of evidence is the independent institutional corroboration in which the concerned authority confirmed that the poaching hotspots align with their understanding of the poaching activity locations in the university. Although informal, this agreement provided an external consistency check independent of the original respondents.

The practical result of the integration appears at the deployment. Blanket coverage of 2,000 acre campus is infeasible. The community-driven hotspot map acts as the prior analysis of where the IoT system needs to be installed, which in this study are the places with low-population, concealed, infrastructure-isolated zones. It directly informed the fog-based system (Siyanwal et al., 2025) for UoH (low-power motion units triggering Raspberry Pi 4 B nodes at a 5-s end-to-end latency). The integration thus closes the loop in which community knowledge resolves where to look, informatics validates those locations, and the IoT layer operationalizes them under real energy and cost constraints.

7 Conclusion

In this paper, we propose a five-stage geospatial data pipeline for constructing wildlife poaching hotspots from community knowledge in contexts where no historical incident data are available. The pipeline explicitly targets cold-start scenarios, which are largely absent from previous work. To convert community knowledge into ground truth labels, we calibrated a ternary (1, 0, −1) response scale for non-expert informants, where the “-1” category encodes explicit epistemic uncertainty. Descriptive analyses showed that community-identified poaching areas at the University of Hyderabad (UoH) are both spatially and ecologically coherent, and clearly distinguishable from non-poaching areas. These poaching zones are predominantly located inside the campus boundary, in regions with reported human population equal to zero, dense concealing canopy cover, greater distances from built infrastructure, and are mainly associated with evergreen broadleaf forest patches and farmlands, located farther from buildings and offices. Insights from this work motivated the design of an energy-efficient wildlife monitoring system for UoH. We further validated the pipeline outputs using machine learning models; among them, a KNN-based model achieved the highest F1 score of 0.96, indicating that poaching hotspots are well separated from non-poaching locations. The results were also informally shared with the relevant authorities (who did not participate as respondents), and they confirmed that the identified hotspots are consistent with their own understanding of poaching locations on campus. Overall, the pipeline is intended to be broadly applicable to data-scarce conservation settings where informal community knowledge serves as the primary source of spatial ground truth.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. While preparing this manuscript, the authors utilized Grammarly and Overleaf’s built-in AI tools to enhance the text’s readability, grammar, and English language style. Following the use of these tools, the authors carefully reviewed and revised the manuscript as necessary and accept full responsibility for the final scientific content.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/friot.2026.1893789/full#supplementary-material

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Supplementary material Supplementary material at the source

Footnotes

Edited by: Khalid Elgazzar, Ontario Tech University, Canada

Reviewed by: Clara Inés Caro Caro, Universidad de los Llanos, Colombia

Ferryati Masitoh, State University of Malang, Indonesia

Footnotes

www.openstreetmap.org/

qgis.org

archive.pfbc-cbfp.org/news_en/items/unep-interpol-enen.html

smartconservationtools.org

www.thehindu.com/news/cities/Hyderabad/Poaching-rampant-on-UoH-campus/article13982649.ece

www.worldpop.org

proba-v.vgt.vito.be/en/product-types/c2/level-3toc

earthengine.google.com

earthexplorer.usgs.gov

https://plugins.qgis.org/plugins/NNJoin/