Globally, wetlands provide important ecosystem services and are critical to supporting wildlife and biodiversity. Anthropogenic disturbances, such as road construction, have a negative impact on wetland conditions and have dramatically reduced their number worldwide by more than 50%. In response to the damage caused by road construction, the Washington State Department of Transportation (WSDOT) mitigates the consequent reduction of functions and the loss of wetlands through restoration efforts, including the monitoring and eradication of invasive vegetation such as reed canary grass (RCG). WSDOT currently maps and monitors invasive species on the ground, which is challenging as they are hard to access due to inundation and dense vegetation.
Compared to field survey methods, drones have the potential to quickly and safely survey large areas, reducing human effort and cost. By focusing on a single mitigation wetland site, we investigated the use of drones as an effective tool to accurately survey RCG. We flew two drones under different light conditions, at ~37 m altitude, collecting data with three different sensors: two built-in 3-band (red, green, blue (RGB)) cameras and an add-on 5-band (RGB, red-edge, near-infrared (NIR)) camera. We used object-based image analysis (OBIA) with random tree machine learning (ML) classifier to create classified maps of RCG cover and test the accuracy of the maps using visual interpretation and confusion matrices.
Results/Conclusions
The data collected with the RGB sensor (1.5 cm/pixel resolution) on an overcast day generated the best results, having an overall map accuracy (OMA) of 88%, an omission error (OE) of 4% and a commission error (CE) of 8%. Based on a confidence level of 80%, we estimated RCG cover to be between 65.6 - 67.1% of the site's area.
Additionally, we considered opportunities and limitations of using drones and OBIA as tools to survey and map invasive species respectively, by comparing drone survey to field survey results. Moreover, we highlight the factors ecologists and natural resource managers must consider when using drones for wetland monitoring such as light condition, resolution (spectral, spatial, temporal), and vegetation phenology, phenotypic characteristics, and composition. In conclusion, we identified future areas of research that include testing the repeatability of these methods at additional wetland sites and increasing the suitability, number, and timing of the field data in support of this work.