Building a representative UAV RGB reference dataset for national-scale satellite mapping of invasive goldenrods (Solidago spp.): an efficient workflow and accuracy drivers
Summary
No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.
Summarise with VIRAT Free to try with your Science Catalogs account.Abstract
Reliable wall-to-wall mapping of invasive plants from satellite imagery depends on representative reference data and transparent quality control. Here we present a nationally distributed UAV RGB reference dataset and an end-to-end workflow designed to support national-scale satellite mapping of invasive goldenrods (Solidago spp.) in Poland. During the peak flowering period (August–September 2024), we acquired 79 UAV orthomosaics across environmentally diverse landscapes. Expert interpreters delineated reference polygons for Solidago and a background class (other vegetation/land cover) and performed independent visual verification to ensure consistent labelling and to identify recurring sources of confusion. For each orthomosaic we derived a feature stack from RGB imagery (spectral indices and texture metrics) and conducted Random Forest classification, with performance evaluated using class-wise user’s accuracy (UA), producer’s accuracy (PA), and F1-score. Across sites, Solidago classification showed greater variability than the background class and was frequently limited by commission errors, as indicated by comparatively low UA relative to PA. A structured review of spatial error patterns, combined with site-level modelling of logit-transformed accuracy metrics, revealed repeatable operational drivers of false positives. Precision (UA Solidago) decreased significantly under stronger direct-sun conditions (higher sun-visibility scores), consistent with increased within-scene contrast, harsh micro-shadows, and occasional glare or specular highlights, which can mimic flowering signals in RGB orthomosaics. Disturbance (mowing) was associated with reduced Solidago performance (lower F1 and marginally lower UA), reflecting increased heterogeneity, patch fragmentation, and transitional vegetation states. Deep shadows and woody/shrub contexts showed weaker negative trends, reinforcing their status as high-risk environments requiring intensified visual quality control. We synthesise best-practice recommendations for efficient UAV reference-data acquisition and targeted visual validation, emphasising strategies to minimise false positives and maximise label consistency across a national sampling frame. This workflow provides a practical foundation for robust training and validation of subsequent wall-to-wall satellite mapping of Solidago density classes.
Fields of Study & Subjects
Remote-Sensing Image Classification
Remote Sensing in Agriculture
Remote Sensing and Land Use
Publication details
- Journal / source
- EarthArXiv (California Digital Library)
- Publication date
- 28 March 2026
- Language
- EN
- Document type
- Preprint