Generative AI Against Poaching: Latent Composite Flow Matching for Wildlife Conservation

Lingkai Kong, Haichuan Wang, Charles A. Emogor, Vincent Börsch-Supan, Lily Xu, Milind Tambe · 2025

Read Generative AI Against Poaching: Latent Composite Flow Matching for Wildlife Conservation by Lingkai Kong, Haichuan Wang, Charles A. Emogor, Vincent Börsch-Supan, Lily Xu, Milind Tambe (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-14342.txt.

Introduction

Lingkai Kong, Haichuan Wang, Charles Emogor, and three colleagues use generative AI to forecast where poachers will strike, in a paper published in two thousand twenty-five. Better forecasts can guide patrols and other steps that protect wildlife.

Older forecasting tools, such as linear models and decision trees, miss the tangled patterns poaching follows across space and time. Flow matching, a newer generative method, is more flexible but hard to train on field data. Two obstacles stand out: many poaching events are never detected, and records are scarce. The team trains the flow in a hidden latent space joined with a model of imperfect detection. They also start the flow from a simple linear forecast rather than from scratch. Together, these choices let the model learn the true pattern of poaching beneath noisy, thin data.

Catalog record

CREATORLingkai Kong, Haichuan Wang, Charles A. Emogor, Vincent Börsch-Supan, Lily Xu, Milind Tambe
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE43 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-14342

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Generative AI Against Poaching: Latent Composite Flow Matching for Wildlife Conservation Lingkai Kong, Haichuan Wang, Charles A. Emogor, Vincent Börsch-Supan, Lily Xu, Milind Tambe Abstract Poaching poses significant threats to wildlife and biodiversity. A valuable step in reducing poaching is to forecast poacher behavior, which can inform patrol planning and other conservation interventions. Existing poaching prediction methods based on linear models or decision trees lack the expressivity to capture complex, nonlinear spatiotemporal patterns. Recent advances in generative modeling, particularly flow matching, offer a more flexible alternative. However, training such models on real-world poaching data faces two central obstacles: imperfect detection of poaching events and limited data. To address imperfect detection, we integrate flow matching with an occupancy-based detection model and train the flow in latent space to infer the underlying occupancy state. To mitigate data scarcity, we adopt a composite flow initialized from a linear-model prediction rather than random noise which is the standard in diffusion models, injecting prior knowledge and improving generalization. Evaluations on datasets from two national parks in Uganda show consistent gains in predictive accuracy. 1 Introduction Poaching is a major, direct driver of wildlife loss. First, the toll on elephants is stark: in Africa, about 100,000 were illegally killed in 2010–2012 for ivory ( Wittemyer et al. 2014 ) . Beyond elephants, pangolins are targeted for their scales (and meat), with about 8.5 million taken for illegal trade in West and Central Africa from 2014–2021 ( University of Oxford 2023 ; UNO 2020 ) . Predators are not spared: in one monitored southern African population (2011–2018), poaching for body parts (including bones) caused 35% of recorded human-caused lion deaths ( Everatt, Kokes, and Lopez Pereira 2019 ) . These facts show the need to direct limited patrol resources to the places where they can best stop poaching, in line with global reports on wildlife trafficking ( UNO 2024 ) . Given the scale of these losses, an effective way to strengthen protection is to forecast poacher behavior and use these predictions to guide patrol planning. Machine learning models have been applied to estimate poaching risk and support data-driven patrol allocation. However, existing approaches typically rely on classical models—such as linear models ( Xu et al. 2021 ) and decision trees ( Thiault et al. 2020 ; Kar et al. 2017 ) —which are unable to capture the complex spatial pattens and the strategic adaptations of poachers in response to patrols. These models often make simplifying assumptions about spatial independence, which can lead to inaccurate risk forecasts and, in turn, less effective patrol deployment. Figure 1: Well-hidden snares and rangers conducting a patrol to locate them. Photos: Uganda Wildlife Authority Recent advances in generative AI, particularly flow matching ( Lipman et al. 2023 ) and diffusion models ( Ho, Jain, and Abbeel 2020 ) , offer powerful distribution-fitting capabilities that can model complex, high-dimensional patterns. These models are well-suited to capturing the nonlinear, strategic, and spatially correlated nature of poaching behavior. However, applying them to real-world poaching datasets presents two key challenges. First, the amount of historical data is typically small, especially in parks with limited monitoring capacity, whereas such generative models are usually data-hungry. Second, detection of poaching events is imperfect: many illegal activities go unnoticed during patrols, and negative labels indicating the absence of poaching are unreliable due to the difficulty of detecting well-hidden signs such as snares in dense vegetation ( Keane, Jones, and Milner-Gulland 2011 ) . To address these challenges, we propose WildFlow , a framework based on latent composite flow matching . Building on occupancy models widely used in ecology ( MacKenzie et al. 2002 ) , we represent poaching activity as a latent state, distinct from the probability of detection during patrols. This separation enables explicit handling of imperfect detection: the latent state evolves through flow matching, while detection is modeled as the observation process in an occupancy framework. To mitigate data scarcity, we further design a composite flow that initializes the generative process from a linear model’s prediction rather than random noise, which is standard in diffusion models. By embedding domain knowledge in the initialization, our method improves generalization and supports effective learning in small-data settings. We summarize our contributions are as follows: (1) We present WildFlow , the first application of generative AI to poaching prediction. (2) We propose a latent composite flow matching framework that jointly addresses the challenges of limited data and imperfect detection in conservation. (3) On real-world datasets from two national parks in Uganda, our method improves AUPR by 7.0% and 10.2% for MFNP and QENP, respectively. We also outline steps toward responsible deployment, including field pilots, and careful rollout. 2 Related Works AI for poaching prediction. Yang et al. 2014 introduced the Protection Assistant for Wildlife Security (PAWS) using a subjective-utility quantal response (SUQR) behavioral model. To explicitly account for imperfect detection, Nguyen et al. 2016 proposed a two-layer Bayesian network with latent variables. However, Kar et al. 2017 later found this approach brittle in practice due to model complexity; an ensemble of decision trees achieved higher accuracy with lower runtime and was validated in a one-month field test. To combine the strengths of prior methods, Gholami et al. 2017 proposed a geo-clustering technique that yields a hybrid of multiple Markov random fields with a bagging ensemble of decision trees, supported by a five-month controlled field test. To further address nonuniform uncertainty stemming from uneven patrol effort, Gholami et al. 2018 ; Xu et al. 2020 trained ensembles of weak learners—decision trees and Gaussian processes—on data stratified by patrol intensity, with deployment via the SMART conservation platform ( SMART 2013 ) . Finally, Xu et al. 2021 used logistic regression to quantify the deterrence effects of ranger patrols on poaching risk. Despite these practical successes, most approaches still rely on classical models with limited capacity to capture high-dimensional, nonlinear spatial patterns; few make use of modern deep learning. Most recently, Kong et al. 2025a applied diffusion models to green security problems. Our work differs in three key ways: (i) Their focus is on robust patrol optimization for general green security, whereas we focus specifically on poaching risk forecasting for conservation. (ii) They do not account for imperfect detection, which is a critical challenge in conservation settings, while we explicitly model the visit-level detection process. (iii) Their features are restricted to historical patrol effort, whereas we incorporate rich environmental covariates. In this richer feature space, diffusion models struggle due to their data-hungry nature. To address this, we propose a composite flow model that improves data efficiency. Generative AI for wildlife conservation In camera‐trap vision, neural generative models have been applied directly to wildlife datasets: CycleGAN variants translate between sensor domains (e.g., visible and near-infrared) and augment rare species to improve few-shot classification ( Gao et al. 2020 ; Zhang et al. 2023 ) . Beyond imagery, generative audio models have been developed to synthesize wildlife vocalizations and strengthen bioacoustic monitoring under data scarcity. Early work explored class-conditional GANs for animal-sound augmentation ( Kim et al. 2023 ) , followed by diffusion-based pipelines that generate birdsong spectrograms to improve classifier accuracy ( Gibbons et al. 2024 ) , and more recently, diffusion models that synthesize

Read the complete text