Towards real-time rainforest gunshot detection using efficient deep learning and acoustic sensor integration

Despite increased surveillance and global collaboration, wildlife poaching remains a major conservation concern. The technological advancement of Autonomous Recording Units (ARUs) has enabled acoustic monitoring over large spatial and temporal scales, creating the possibility of real-time alerting systems for poaching. However, high false-positive rates (FPRs) and large computational needs of existing models have rendered real-time detection of poaching impractical. I propose a lightweight rainforest gunshot detection model suitable for on-device processing and a sensor consensus function to minimize false-positives. The developed model, trained on augmented log-mel spectrograms from Belize, consisted of both 2-D and 1-D depthwise-separable convolutional layers and a Gated-Recurrent layer. High FPRs are a pervasive challenge for gunshot detection systems due to the acoustic similarity of commonplace sounds such as cracking branches. SAIL (Sensor Analysis and Integration Layer) was developed to reduce false-positives by leveraging gunshot propagation properties. Testing through acoustic propagation simulation on a spatially distinct dataset across 1000 iterations showed significant FPR reduction when comparing SAIL probabilities to simple inference. The developed model achieved 0.908 validation F1 at 11MB (964KB serialized), retaining 89% of existing literature performance while reducing model size by over 98%. Altogether, this approach presents a promising approach to cost-effective real-time poaching detection.