Deep Learning Assessment of Ecotourism Effects on Wild Mammal Behaviour

Abstract

Ecotourism is increasing near protected areas with wild mammals, and though it can create a conservation income and awareness, it can simultaneously impact feeding, resting, vigilance, movement, and nocturnal activity patterns of the mammals it is promoting. Traditional methods for evaluating these impacts involve direct field observations and manual analysis of camera trap images, which are both time-consuming and susceptible to observer bias and cannot keep up with the constant flow of data produced by today's reserves. This work presents a deep learning-based assessment tool combining a CNN for mammal recognition and species identification with a CNN-BiLSTM temporal module for behaviour recognition, and integrating the behavioural predictions with tourism-pressure factors. The behavioural monitoring data includes camera-trap imagery and video, acoustic recordings, GPS tracks and tourist-activity logs, all gathered within both the tourist and control areas. The framework is assessed by means of accuracy, precision, recall, F1-measure, AUC and confusion matrix analysis and compared with CNN, ResNet, LSTM and traditional machine-learning classifiers. The framework is supposed to be able to identify the measurable increase of vigilance and avoidance behaviour, and a shift to become more nocturnal in high-tourism areas compared to control areas, and have high performance in low light and partial occlusion. The framework provides a method for increasing tourists to a protected area while minimizing disturbance, while maintaining the ecotourism benefits, by correlating quantitative behavioural data with specific tourism variables and providing a non-invasive tool that can be interpreted by protected-area managers