Multivariate Deep Learning for Bird Species Sighting Forecasting: A Case Study in Las Lagunas De La Mata y Torrevieja Natural Park

Abstract

Accurate prediction of bird species sightings is vital for enhancing tourism planning and supporting biodiversity management, allowing visitors to natural parks to optimize their schedules for birdwatching experiences. This study focuses on the Natural Park of Las Lagunas de la Mata y Torrevieja (Alicante, Spain), where advanced artificial intelligence (AI) models were developed to forecast the presence of multiple bird species simultaneously through a multivariate modeling strategy that integrates shared temporal dynamics and interspecific predictive relationships among species. The analysis revealed that data granularity significantly impacts the performance of predictive models. High-frequency datasets exhibited substantial noise and a high proportion of zero-valued observations, making daily forecasting particularly challenging. In contrast, aggregated datasets with weekly or monthly intervals exhibited clearer seasonal patterns and reduced noise, favoring recurrent architectures such as LSTM and CNNLSTM, whose relative performance varied depending on the temporal granularity and evaluation metric. The final configurations were evaluated through recursive multi-step forecasts from multiple temporal origins and random seeds, and a target-only comparison showed that using all selected species as input reduced RMSE and MAE compared with models trained only on the target species. Additionally, a web augmenter was implemented to enhance the Wikipedia page for this natural park, offering interactive visualizations such as dynamic maps displaying species-specific bird sighting predictions. This approach facilitates tourism planning and illustrates the potential of integrating AI-driven multivariate forecasting with web augmentation technologies to support biodiversity monitoring and public engagement.