Cross-Model Confusion Mapping: False Alarm Minimisation in Possum Call Detection Using Deep Learning

Abstract The common brushtail possum is a major invasive species in New Zealand, damaging native ecosystems and agricultural systems and is a target of the national Predator Free 2050 eradication programme. Reliable, low-cost monitoring is critical to this programme, and acoustic detection offers a scalable alternative to traditional trapping and camera-based methods. The bioacoustics approach to automated brushtail possum detection using deep learning models suffers from excessive false alarms. At the same time, the high computational demands of the most accurate models make them unsuitable for resource-limited edge devices. In this paper, a targeted hard-negative mining method, Cross-Model Confusion Mapping (CMCM), is proposed to overcome these problems. It uses the error profile of a pre-trained model to identify bird calls most frequently misclassified as possum vocalisations. These bird calls are then relabelled and added to the training data. Three CMCM audiosets were generated and compared with size-matched sets assembled through untargeted sampling. CNNs with 3, 5, 7, 10, and 13 layers were trained using identical augmentation pipelines. Experiments showed that CMCM-trained models maintain accuracy while significantly reducing false alarms. In particular, the 10-layer CMCM-trained CNN achieved the most reliable detection with zero false positives. The inference times are orders of magnitude faster, with substantially lower computational demand, than the state-of-the-art Audio Spectrogram Transformer. These findings suggest that CMCM enables accurate, low-latency possum detection suitable for real-time deployment on edge devices. The method is generalisable to other bioacoustic detection tasks where false positives arise from acoustically similar non-target species.