Invasive Species Monitoring with Fine-Resolution Multispectral sUAS-acquired Imagery and a Deep Neural Network Classifier

Abstract:

The invasive plant Phragmites australis negatively impacts native wetland plants and migratory bird habitats in the Great Salt Lake ecosystem. The Utah Department of Natural Resources is applying treatments to reduce Phragmites cover and requires an efficient monitoring program for effective remediation. Due to the difficulty of navigating the wetland environment, lower cost, and better image quality than satellite or traditional airborne products, sUAS acquired multispectral imagery is well-suited for mapping Phragmites in this complex wetland ecosystem. In July 2020, 7.5 cm resolution 5-band imagery was captured over 3.2 km2 in the Howard Slough Waterfowl Management Area, Utah, using a Parrot DISCO fixed-wing UAS with a Red-Edge-MX sensor. Phragmites and native vegetation cover were classified using a convolutional neural network with Keras and a Tensorflow backend. In a test campaign, four orthoimages covering about 1.6 km2 were classified, and kappa index values ranged from 0.77-0.87, indicating substantial agreement. These classifications are used for the evaluation of remediation programs and for planning future campaigns. The imagery was also resampled to coarser resolutions and classified to simulate results for higher flight altitudes that reduce flight time, and for comparison to resolutions equivalent to airborne and satellite products. Our results demonstrate the pairing of sUAS and deep learning neural networks as a valuable option for natural resource monitoring and invasive species mitigation.