Investigating hunting in a protected area in Southeast Asia using passive acoustic monitoring with mobile smartphones and deep learning

• Hunting poses a significant threat to wildlife. • We recorded gunshots in Vietnam and used transfer learning for automated detection in the ‘torch for R’ ecosystem. • We found that transfer learning architectures led to acceptable performance. Hunting is one of the most serious threats to wildlife populations. Guns can be used to hunt both terrestrial and arboreal species. While many methods and techniques have been developed to monitor wildlife populations, few techniques have been developed that can meet the demand needed to monitor threats to wildlife populations and support intervention activities; this is particularly true in Southeast Asia. Here, we used smartphones to record gunshots at 64 sites that were systematically spaced in Chu Mom Ray National Park, Vietnam from July to November 2019. Our specific goals were to: 1) quantify temporal and spatial patterns in gunshots in the national park; 2) investigate the correlation between gunshots and seven environmental variables including forest quality, elevation, and distance to the nearest village or ranger post; and 3) compare the performance of three convolutional neural network (CNN) architectures for automatically detecting gunshots. We manually annotated spectrograms of gunshots using Raven Pro 1.6. Using the manual detection approach, we identified 115 gunshots at 30 sites. The number of gunshots recorded at each recording site over two days varied from 0 to 11. On average, there were about 0.93 gunshots recorded per day per recording site. Hunting activity showed a strong temporal trend. The number of gunshots detected increased gradually from early morning, reaching the peak at noon, with the most gunshots recorded from 10:00 to 14:00 local time, equivalent to 0.18 gunshots per hour per site. No gunshots were detected from 22:00 to 04:00 in the morning. There were no strong relationships between the number of gunshots and all environmental variables, with all the correlation coefficients smaller than 0.3. Comparing three CNN architectures — AlexNet, VGG16, and ResNet18— implemented in the ‘torch for R’ ecosystem, we found that both AlexNet and VGG16 architectures led to acceptable performance for automated detection (F1 score > 0.80), but the ResNet18 architecture did not perform well for this task. We found low generalizability, as the models had relatively low performance on an open dataset from Belize (F1 score ∼ 0.52). To implement effective patrols, protected-area managers should make a hunting-risk map. The effort required for this task can be substantially reduced by using a combination of PAM and automated gunshot detection, however these approaches still require creation of a site-specific training and test dataset, along with manual verification of the detections.