Deep RNN with Pseudo Loss Objective for Forecasting Stop-over Decisions of Wild Migratory Birds

Forecasting stop-over decisions and mapping the stop-over sites of wild migratory birds is fast becoming important in light of recent developments affecting global health. Migratory wild birds stop at sites with access to food resources so they can rest before continuing with their journey. Unfortunately, these sites offer opportunities for these birds to spread pathogens and viruses by interacting with the ecosystem. While previous work has focused on predicting stop-over sites using historical information, we emphasize that this is not useful for any planning efforts by health authorities and instead offer a new perspective by proposing an approach that can forecast the duration of stop-over. In this work, first we cast this problem as a classification task and show how pseudo labels and losses in a Bi-directional recurrent neural network can help improve performance given the presence of significantly underrepresented class. We use dataset of Turkey vulture (avian pox vector) movement over several years for the forecasting task where we compare our approach with a variety of baselines and show that it outperforms them. We also use this dataset and the White Fronted Geese (avian flu vector) movement dataset to understand the nature of the habitats used for stop-over using a publicly available model pre-trained on more than half a million land cover images. By knowing the preferred stop-over habitats and the time spent in and between stop-overs using our model, we can help relevant authorities come up with efficient intervention measures.