Suggesting an Efficient Deep Learning Architecture for Coastal Wetland Land Cover Mapping with UAS Imagery

Deep learning has already been shown to be a powerful state-of-the-art technique for many applications including image understanding and remote sensing (RS) image analysis. Unmanned aerial systems (UAS) offer a viable and economical alternative to conventional sensors and act as a platform for acquiring high spatial and high temporal resolution data with high operational flexibility. Coastal wetlands are among the most complex yet important ecosystems for land cover mapping due to the fact that land cover targets show high intra-class and low inter-class variances. In recent years, several deep learning architectures have been proposed for high-performance pixel-wise image labeling. In this experiment, a new dataset for coastal wetland segmentation containing 10,000 image patches with a resolution of 960*720 pixels is provided. The most recent deep learning architectures proposed for semantic segmentation are described and their performances on a manually labelled UAS dataset are investigated. Our experiment has found and suggested an easy to fine-tune deep learning architecture for accurate land cover mapping applications which needs less training samples and remarkably fewer training epochs.