Semantic Segmentation with Multispectral Satellite Images of Waterfowl Habitat


 Migratory waterfowl (i.e., ducks, geese, and swans) management relies on landscape bioenergetic models to inform on-the-ground habitat conditions conservation practices. Therefore, planners rely accurate predictions of wetland habitats for at regional scales. Unharvested flooded corn is a popular tool public private lands that greatly increases landscape-level energy compared other wetlands; thus, are particularly sensitive these features. Despite their importance planning implementation, the abundance distribution unharvested fields across North America unknown. Furthermore, training data difficult collect challenging given unique attributes discreteness lens. Advances in multispectral imagery deep learning algorithms may enable continuous autonomous detection we conducted modeling experiments using West Tennessee collected from Sentinel-2 satellite missions. We performed several individual band combination composites and/or vegetation indices identify optimal bands MRUNET architectures. subsequently used 3 ensemble important networks. found use was necessary although CIR composite OSAVI index improved precision, 12-band increased recall, metric were most interested in. Moreover, all ensembles exhibited poor performance. Here, present results our initial suggest future exercises including temporal image stacking multi-modal recurrent neural network architectures.