Predicting bird species detection in avian ecology using machine learning models, environmental factors, and tree characteristics

Abstract

Birds are a critical unit of the ecological system but determining what attracts or discourages them from a given home range remains a challenge. This study was carried out to fill this gap. Ecological data were collected from the field using survey methods, then developing presence–absence models for bird species using advanced machine learning techniques. Stratified random sampling was adopted to ensure that all the four habitats (degraded areas, wetlands, forestland, and bare ground) were represented. Line transect techniques were used to count bird species during data collection and bird count was carried out every fortnight. Environmental factors such as air quality (ozone, PM10, PM2.5, TVOC) and temperature were measured using air quality monitor, and tree height was measured using clinometer. Different models such as logistic regression, multilayer perceptron (MLP), decision tree, support vector machine, ensemble learning (using voting classifier), random forest and deep neural networks (DNN with L2 regularization) were developed to predict the presence or absence of bird species. The models were trained, and each model was evaluated using reliable metrics (accuracy, AUC, Brier score, Log loss) to determine their predictive abilities as well as calibration on external ecological datasets. The results were cross-examined with care, allotting more curiosity to the applied value, strengths, and weaknesses linked to each model in real-world setting, with DNN and random forest showing first-rate predictive accuracy of 94.3% and 86.3% when evaluated using independent datasets. Feature importance was analyzed and species activity, time of the day, and some pollutants were influential predictors of probability of detection of birds. These models differ from native statistical technics because they provide more information about the outcome which could be helpful in ecological conservation efforts not just because they are easy to implement complex analyses. Machine learning presents a more pragmatic and impressive tool for real-time biodiversity monitoring and steering data-driven conservation decision-making.