Detection of trade in products derived from threatened species using machine learning and a smartphone application

Ritwik Kulkarni; WU Hanqin; Enrico Di Minin

cs.CVSep 8, 2025

Detection of trade in products derived from threatened species using machine learning and a smartphone application

Authors: Ritwik Kulkarni, WU Hanqin, Enrico Di Minin

Organizations: Helsinki Lab of Interdisciplinary Conservation Science, Department of Geosciences and Geography, FI-00014, University of Helsinki, Finland · International Fund for Animal Welfare (IFAW) · School of Life Sciences, University of KwaZulu-Natal, Durban 4041, South Africa

Abstract

Illegal and unsustainable wildlife trade remains a major conservation challenge and is increasingly shifting to online marketplaces where monitoring is difficult. To address this gap, we used real-world images from online platforms, enforcement seizures, and market inspections to develop a deep-learning object detection framework to identify products derived from three highly traded species: elephant, pangolin, and tiger. Our dataset of 39,941 images included both wildlife products and visually similar non-wildlife items, reflecting real enforcement complexity. We used cross-entropy and asymmetric focal loss functions to train Faster Region-based convolutional neural network (Faster R-CNN) models and compared species-specific and multispecies training strategies. The best-performing model, trained with cross-entropy on all species combined, achieved an overall accuracy of 84.2%, with high performance for pangolin (93.6%) and tiger (90.3%) and moderate performance for elephant (71.1%). Although asymmetric focal loss increased recall, it substantially reduced precision. We deployed the optimized model in a prototype smartphone application, potentially usable in real-world settings for monitoring online images. With an independent test set, the application achieved 91.3% accuracy and an F-score of 0.86. These results demonstrate the potential of object detection models to support scalable, real-time monitoring of wildlife trade and enable the rapid triage of suspicious posts, regulatory compliance, and conservation decision-making.

Figures & tables

Figure 1: Machine learning based object recognition models were trained on images of elephant ivory and skins, pangolin scales and claws, and tiger skins and bones. The optimized model achieved high detection accuracy and was hosted on the cloud to be integrated into a smartphone application. This enabled real-time identification of potentially illegal products. This pipeline supports both online monitoring of wildlife trade and on-the-ground enforcement by authorities.

elephant
ambiguous mammoth	5409
antique ivory	414
elephant skin	22
ivory	6893
non-widlife	12890
sub-total	25,628

Table 1: Counts of labeled images for elephant (ivory, skin, antique ivory, and ambiguous mammoth), pangolin (raw, parched, and crafted scales, claws), and tiger (bones, claws, fangs, and other products), along with corresponding non-wildlife categories. The total number of images were 39,941. These datasets formed the basis for training and evaluation of object recognition models to detect wildlife products in digital content.

Figure 2: Screenshot of the app depicting the prediction of detected ivory, crafted pangolin scale and ambiguous tiger fang products. The app also has the ability to collect feedback on the results of the prediction.

Figure 3: Performance comparison of multi-species model trained across all three species. Bar chart illustrating model performance using two loss functions; cross-entropy loss and asymmetric focal loss, on a combined dataset of elephant, tiger, pangolin, and non-wildlife categories. Metrics include mean average precision at 50% IoU (mAP_50), mean average recall at 100 detections (mAR_100), category-wise mAP and mAR, and overall accuracy. Under cross-entropy loss, the model achieved higher performance, with overall mAP_50 of 0.60, mAR_100 of 0.56, and species-level accuracies of 71.1% (elephant), 90.3% (tiger), 93.6% (pangolin), and 82.1% (non-wildlife). In contrast, asymmetric focal loss produced lower accuracies, 30.7%, 9.2%, 3.3%, and 51.7% for elephant, tiger, pangolin, and non-wildlife, respectively, despite slightly higher recall for some categories. The visualization highlights the relative robustness of cross-entropy loss for detecting wildlife products across diverse taxa and underscores the challenge of optimizing precision–recall trade-offs when developing a single model for multi-species detection.

Figure 4: Bar charts summarize the performance of individual models trained on single-species datasets under two loss functions; cross-entropy loss and asymmetric focal loss. Each row panel represents one species (elephant, pangolin, and tiger), with results shown for species products versus non-wildlife images. Performance metrics include mean average precision at 50% IoU (mAP_50), mean average recall at 100 detections (mAR_100), category-wise mAP and mAR, and classification accuracy. Across species, cross-entropy consistently produced higher accuracies—65.0% for elephant products, 70.6% for pangolin products, and 82.7% for tiger products, compared to substantially lower accuracies under asymmetric focal loss (50.0%, 38.1%, and 29.9%, respectively). While focal loss achieved comparable or slightly higher recall in some categories, it generally reduced precision and overall accuracy. Multi-species model (Figure 3 ) showed better performance across most of the metrics compared to per species models.

Figure 5: Gradient-weighted Class Activation Mapping (Grad-CAM) outputs for Faster R-CNN models applied to (a) carved elephant ivory, (b) pangolin scales, and (c) tiger claws. Heatmaps highlight key discriminative features: intricate carvings and textures in ivory, localized scale patterns in pangolin, and sharp curved edges in tiger claws. These activations illustrate how the model identifies species-specific visual cues for classification.

Species	TP	FP	FN	Precision	Recall	F-Score	Sensitivity
Elephant	38	15	1	0.72	0.97	0.83	97.4%
Tiger	35	2	4	0.95	0.89	0.92	89.7%
Pangolin	38	9	1	0.80	0.97	0.88	97.4%
Non Wildlife	15	3	6	0.83	0.71	0.76	71.4%
Overall	126	29	12	0.81	0.91	0.86	91.3%

Table 2: Real-world performance of the deployed smartphone application. Performance metrics of the trained model when served through a mobile app and tested on live images of elephant, tiger, and pangolin products, as well as non-wildlife items. TP stands for True Positive, FP - False Positive, and FN False Negative. The app achieved high sensitivity for elephant (97.4%), pangolin (97.4%), and tiger (89.7%) products, with lower performance on non-wildlife images (71.4%). Overall sensitivity (same as total accuracy) was 91.3%, demonstrating the model’s effectiveness for real-time detection in practical settings. In the same order F-score was 0.83, 0.92, 0.88, 0.76 and 0.88.

Object Detection

Object detection locates and classifies objects in images or video, balancing localization accuracy, robustness, and inference speed.

See all 624 papers

Automated identification of ivory carvings using machine learning: A proof-of-concept approach to detecting the illegal wildlife trade
Oct 23, 2026Andrius Tamosiunas, Simone Mozzon, David Bacon +2Image Classification
Institute of Theoretical Astrophysics, University of Oslo, PO Box 1029, Blindern 0315, Norway · CERCA/ISO, Department of Physics, Case Western Reserve University, 10900 Euclid Avenue, Cleveland, OH, 44106, USA · Institute of Cosmology and Gravitation, University of Portsmouth, Dennis Sciama Building, Burnaby Road, Portsmouth, PO1 3FX, UK +2
Democratising Camera Trap AI: An Open-Source Model for Detecting UK Mammals
Jun 9, 2026Paul Fergus, Philip Stephens, Russell A. Hill +8You Only Look Once Object Detection
1Liverpool John Moores University, Byrom Street, Liverpool, L3 3AF, UK. · 2Durham University, Stockton Road, Durham, DH1 3LE. · 3MammalWeb, c/o Durham Wildlife Trust, Houghton-le-Spring, DH4 6PU +6
A unified three-branch deep learning framework for automated detection and counting of wildlife in aerial imagery
Oct 12, 2026Zeyu Xu, Tiejun Wang, Andrew Skidmore +5Image Matching Wildlife Monitoring
School of Geospatial Engineering and Science, Sun Yat-Sen University, Zhuhai, China · Faculty of Geo-Information Science and Earth Observation, University of Twente, Enschede, The Netherlands · Boma and Badingilo National Parks, African Parks, Juba, South Sudan +3

Oct 23, 2026J/K move · Enter open · S save

Automated identification of ivory carvings using machine learning: A proof-of-concept approach to detecting the illegal wildlife trade

Andrius Tamosiunas, Simone Mozzon, David Bacon, Caroline Cox+1

Institute of Theoretical Astrophysics, University of Oslo, PO Box 1029, Blindern 0315, Norway · CERCA/ISO, Department of Physics, Case Western Reserve University, 10900 Euclid Avenue, Cleveland, OH, 44106, USA · Institute of Cosmology and Gravitation, University of Portsmouth, Dennis Sciama Building, Burnaby Road, Portsmouth, PO1 3FX, UK+2

Image Classification

Open paper AskPDF arXiv

Despite a ban on the international trading of recently poached ivory, elephants are still killed for their tusks. Ivory and its products are mainly sold in online marketplaces. However, due to the huge volume and variety of adverts and the similarity of ivory with other materials that can be legally sold, it is hard for enforcement officials to identify illicit trading and prosecute sellers. Existing techniques for identifying elephant ivory, such as Fourier-transform infrared spectroscopy and microscopy, can reliably distinguish ivory from other materials, but require physical access to the item and cannot be applied to photographs alone. As a result, these methods are unsuitable for screening the large volume of images posted in online marketplaces, leaving a gap for automated, image-based identification tools. In this proof of concept work, we investigate the potential of machine learning techniques for image analysis to automatically distinguish ivory objects from bone. We employ a semantic segmentation framework using a U-Net architecture with a ResNet34 backbone, leveraging transfer learning on an augmented dataset of 541 original images. Using a dataset of images of ivory and items resembling carved ivory, several models were trialled. On a distinct test set, the classifier achieved an Area Under the Curve (AUC) of 0.81 and an overall classification accuracy of 81% on confident samples. The paper presents the best-performing model as a proof of concept for developing a system that can effectively identify ivory and its products when indicative features of elephant ivory are present in an image. This work provides initial steps towards a machine learning-based framework designed to advise law-enforcement in detecting illegal elephant ivory trade.

Abstract

Object Detection

Explore similar work

Automated identification of ivory carvings using machine learning: A proof-of-concept approach to detecting the illegal wildlife trade

Democratising Camera Trap AI: An Open-Source Model for Detecting UK Mammals

A unified three-branch deep learning framework for automated detection and counting of wildlife in aerial imagery