Effective management of invasive species requires rapid detection and dynamic monitoring. Remote sensing offers an efficient alternative to field surveys for invasive plants; however, distinguishing individual plant species can be challenging especially over geographic scales. Satellite imagery is the most practical source of data for developing predictive models over landscapes, but spatial resolution and spectral information can be limiting. We applied deep learning neural networks to two types of satellite imagery in an effort to detect the invasive plant, leafy spurge (Euphorbia virgata), across a heterogeneous landscape in Minnesota, USA. We first developed convolutional neural networks (CNNs) with static imagery from Worldview-2 and Planetscope satellites. Worldview-2 imagery has high spatial and spectral resolution, but images are not routinely taken in space or time. By contrast, Planetscope imagery has lower spatial and spectral resolution, but images are taken daily across Earth. Second, we modified the CNN for Planetscope to include a long short-term memory (LSTM) layer that leverages information on growth and flowering phenology from a time series of images.
Results/Conclusions
The model using high-resolution Worldivew-2 imagery had 96.1% accuracy in detecting leafy spurge, whereas the latter model using Planetscope imagery had 89.9% accuracy. The detection accuracy of the Planetscope LSTM model was 96.3%, which is on par with the high-resolution, Worldview-2 model. Across models, most false positive errors occurred near true populations, indicating that these errors are not consequential for management. We identified that early and mid-season phenological periods in the Planetscope time series were key to predicting leafy spurge. Additionally, green, red-edge, and near-infrared spectral bands were important for differentiating leafy spurge from other vegetation. These findings suggest that deep learning models can accurately identify individual species over complex landscapes even with satellite imagery of modest spatial and spectral resolution if a temporal series of images is incorporated. Our results will help inform future management efforts using remote sensing to identify invasive plants, especially across large-scale, remote, and data-sparse areas.
