Object- and pixel-oriented species mapping in mangrove forests using multisource spaceborne imagery and deep learning

Mapping tree species is critical in mangrove forests to assist in monitoring and conservation operations. Here, we present a multi-task deep learning framework that performs accurate delineation of tree crowns and classification of species in mangroves using very-high-resolution spaceborne imagery. By matching the two outputs, a final mapping of species at the tree level is achieved. Our framework also performs well on high- and medium-resolution imagery, including the newly-operating hyperspectral missions.