Invasive species are a significant management concern in grasslands globally. Remote sensing observations from various platforms have been used to monitor invasive species and reduce the cost associated with field data collection. These monitoring approaches often rely on remote sensing data from only one or a few time points, with little attention to the phenology (life cycle events) of invasive plants. Yet grasslands exhibit high temporal variability, making it necessary to develop scalable approaches that map biological invasions across the growing season. Here, our goal is to develop a remote sensing-based, ecologically informed approach for detecting invasive plants in heterogeneous grasslands. As a case study, we mapped Lespedeza cuneata ( L. cuneata ), an aggressive invasive legume at the Tallgrass Prairie Preserve, Oklahoma, U.S. using time series of PlanetScope imagery. Specifically, we used a combination of vegetation indices and land surface phenology metrics as inputs for knowledge-guided machine learning models, including Random Forest, Support Vector Machine, and Artificial Neural Network. Classification models were then developed and validated for L. cuneata across growing seasons and three phenological phases (green-up, mid-season, and senescence). Results showed that land surface phenology metrics of L. cuneata patches differed significantly from cooccurring native species, and incorporating these variables consistently improved classification performance (78–83% overall accuracy) compared to models using only vegetation indices (54–75%). Our findings demonstrate that integrating phenological dynamics with fine resolution remote sensing data in a knowledge-guided machine learning framework substantially enhances invasive species detection in grasslands, supporting more ecologically informed land management across spatial scales. • Integrating phenology metrics yields consistently high accuracy in detecting invasive plants • Mid-season and senescence periods show greater detection accuracy • Phenology metrics perform better than spectral indices alone as predictors
Arora et al. (Sun,) studied this question.