Fully Convolutional Neural Network for Land Cover Mapping In A Coastal Wetland with Hyperspatial UAS Imagery

Coastal wetlands are among the most productive ecosystems in the world providing numerous valuable services for people and wildlife including recreation, fish and wildlife protection, sediment control and flood prevention. Changes in land cover/use have great impact on the functionality and productivity of wetlands. Therefore, accurate and fast monitoring of land cover/use allow policy makers and wetland users to devise and implement policies and management practices to lessen the side effects of any potential change on wetlands. In this study, we apply deep convolutional neural network (DCNN) to infer complex wetland classes in a semantic segmentation task on hyperspatial resolution UAS images. We examine the capacity of the well-known FCN-VGG architecture for wetland mapping. Results illustrate that fine-tuning the network using transfer learning with UAS hyper-resolution images provide state-of-the art semantic segmentation accuracy of 89.98%.