Deep Semantic Segmentation of Trees Using Multispectral Images

Forests can be efficiently monitored by automatic semantic segmentation of trees using satellite and/or aerial images. Still, several challenges make the problem difficult, including varying spectral signature different trees, lack sufficient labelled data, and geometrical occlusions. In this paper, we address tree multispectral imagery. While carry out large-scale experiments on deep learning architectures various input combinations, also attempt to explore whether hand-crafted vegetation indices improve performance models in trees. Our include benchmarking a variety remote sensing image sets, architectures, bands as inputs, number indices. From our experiments, draw useful conclusions. One particularly important conclusion is that, with no additional computation burden, combining categories indices, such NVDI, ARVI, SAVI, within single three-channel input, state-of-the-art accuracy improved under certain conditions, compared high-resolution visible nearinfrared input.