Detection and classification of invasive Callery pear from multisource satellite imageries in Google Earth Engine

Assessing the spatial distribution of invasive species is a critical component for establishing baselines to examine the rate of spread, documenting key areas where negative impacts on ecosystems can be mitigated, and developing sustainable management strategies. Callery pear ( Pyrus calleryana Decne.; PC) is a rapidly spreading invasive woody plant species in the eastern United States (U.S.). The incursion of these wild-type trees into urban, peri-urban, and rural landscapes, which have escaped from transplanted clonal cultivars, poses complex management challenges for land managers and hampers ecosystem function by competing with native plant species and altering the forest environment. We integrated winter season multispectral (Sentinel-2) and radar (Sentinel-1 and L-band) imageries with static terrain imagery to map PC distribution in four southeastern U.S. states. From these imageries, we derived spectral, textural, and elevational indices and used them with field-collected PC locations for training random forest (RF) and support vector machine (SVM) classifiers in Google Earth Engine. We also created four scenarios to sequentially add and determine the usefulness of input data in classifying PC. The scenario comprising all the imageries plus their derived indices was most accurate for RF (Accuracy [Low CI, Upper CI] = 92.6% [90.62, 94.34]) and SVM (89.6% [87.45, 92.00]), with the terrain and L-band radar indices identified as the most important inputs for discriminating PC from other classes. Accuracy increased by 10.5% for RF and 5.2% for SVM for the best scenario compared to using bands and derived indices from Sentinel-2 imagery alone. Our final classification model identified a relatively greater PC spread in the northeastern part of our study area, eastern Tennessee. Combining L-band radar in classification scenarios enhanced PC classification. Our approach demonstrates the utility of several remote sensing images in mapping and monitoring the distribution of invasive plant species on a large scale. • Practical approach to overcome challenges in large-scale invasive species mapping using Google Earth Engine. • Integrating multispectral and radar imagery with terrain can enhance Pyrus calleryana maps at the regional scale. • Indices derived from L-band imagery are superior compared to other indices for Pyrus calleryana detection. • Regions from eastern Tennessee and upland South Carolina were extensively infested by Pyrus calleryana .