Fusion of Lidar and Aerial Imagery to Map Wetlands and Channels via Deep Convolutional Neural Network

Abstract

Wetlands and channels provide significant ecological and societal services. Mapping their locations and types at high resolution remains a challenge for scientific communities and regulatory agencies. In the U.S.A., the National Wetland Inventory (NWI) provides the best nationwide wetland maps, but the NWI for some locations has not been updated for up to four decades and includes omission errors. To address these problems, we developed a deep learning framework for identifying wetlands and channels from lidar point clouds and 1 m aerial images. Both deep learning and terrain analysis were applied to classify wetlands and identify channels. The deep learning classifier was a convolutional neural network that utilized both image color information and lidar-derived canopy height. When tested on a 94 km2 Ohio watershed, the method achieved a classification accuracy of 85.6%. Compared with the NWI maps, the results included 76% more forested wetlands by area and 168% more channels by length. It was also found that only leaf-off images were useful for detecting forested wetlands and typical commission errors (i.e., false positives) were attributed to tree shadows. The study demonstrates the advantages of combining lidar structural and aerial spectral information over the use of each alone and exemplifies the utility of deep learning as an effective means to leverage open-source data for high resolution mapping of wetlands.