Evaluating the Efficacy of Sentinel-2 versus Aerial Imagery in Serrated Tussock Classification

Invasive species pose major global threats to ecosystems and agriculture. Serrated tussock (\textit{Nassella trichotoma}) is a highly competitive invasive grass species that disrupts native grasslands, reduces pasture productivity, and increases land management costs. In Victoria, Australia, it presents a major challenge due to its aggressive spread and ecological impact. While current ground surveys and subsequent management practices are effective at small scales, they are not feasible for landscape-scale monitoring. Although aerial imagery offers high spatial resolution suitable for detailed classification, its high cost limits scalability. Satellite-based remote sensing provides a more cost-effective and scalable alternative, though often with lower spatial resolution. This study evaluates whether multi-temporal Sentinel-2 imagery, despite its lower spatial resolution, can provide a comparable and cost-effective alternative for landscape-scale monitoring of serrated tussock by leveraging its higher spectral resolution and seasonal phenological information. A total of eleven models have been developed using various combinations of spectral bands, texture features, vegetation indices, and seasonal data. Using a random forest classifier, the best-performing Sentinel-2 model (M76*) has achieved an Overall Accuracy (OA) of 68% and an Overall Kappa (OK) of 0.55, slightly outperforming the best-performing aerial imaging model’s OA of 67% and OK of 0.52 on the same dataset. These findings highlight the potential of multi-seasonal feature-enhanced satellite-based models for scalable invasive species classification.

💡 Research Summary

The paper investigates whether multi‑temporal Sentinel‑2 satellite imagery can match or surpass high‑resolution aerial photography for mapping the invasive grass Serrated tussock (Nassella trichotoma) across a 1,200‑ha grassland in western Greater Melbourne, Victoria, Australia. Ground truth data were collected in 2021‑2022 using a systematic 10‑m radius plot grid (6,879 plots) and classified into four cover classes (None, Low, Medium, High). Sentinel‑2 Level‑2A products (10‑m bands B2‑B8, B8A, B11, B12) were resampled, and four seasonal composites (Autumn, Winter, Spring, Summer) were generated by taking the median of cloud‑free observations. Nine vegetation indices—NDVI, EVI, SAVI, IRECI, NDVI standard deviation, TDVI, NLI, MNLI, and a custom index—were computed for each season, providing a rich spectral‑temporal feature set. In addition, Grey‑Level Co‑occurrence Matrix (GLCM) textures (contrast, dissimilarity, homogeneity, energy, correlation, angular second moment) were extracted per band.

Eleven classification models were built, varying in feature composition: (1) baseline spectral bands plus 14 legacy indices (M17, M24, M41); (2) texture‑only models (M24, M66); (3) models that combine full‑band spectra, multi‑seasonal composites, and the full suite of nine indices (M40, M76*, M64, M20). Principal Component Analysis retained 99.9 % variance before feeding data to a Random Forest classifier (300 trees). Training used 80 % of the plots, with the remaining 20 % for validation; a fixed random seed ensured reproducibility.

Performance was measured with Overall Accuracy (OA), Kappa (OK), class‑wise F1‑score, and Recall. Texture‑only models performed poorly (OA ≈ 34‑36 %). Models that incorporated multi‑seasonal information and vegetation indices achieved markedly higher results. The best model, M76* (full‑band + GLCM + all nine indices across all four seasons), reached OA = 68 % and Kappa = 0.55, slightly exceeding the benchmark aerial‑image model from Pham et al. (OA = 67 %, Kappa = 0.52). Class‑wise, M76* delivered the highest F1 for the “Medium” (0.51) and “High” (0.80) cover classes, and the highest Recall for “Low” (0.84). Comparisons such as M19 (single‑season indices) vs. M76* (four‑season indices) showed OA improvement from 61 % to 68 % and notable gains in medium‑cover detection, confirming the value of phenological information.

Cost analysis highlighted that Sentinel‑2 is freely available with a 5‑day revisit, whereas high‑resolution aerial imagery incurs substantial acquisition and processing expenses. The modest 1 % OA advantage of aerial data does not outweigh its higher cost, making Sentinel‑2 a more scalable solution for regional invasive‑species monitoring.

The study concludes that, despite its coarser spatial resolution, Sentinel‑2 can rival or surpass aerial imagery when enriched with multi‑seasonal composites and a comprehensive set of vegetation indices. The approach improves detection of low‑to‑medium infestations, which are traditionally difficult due to spectral mixing with native grasses. Future work should explore finer temporal granularity (e.g., weekly composites) and deep‑learning time‑series models to further boost medium‑cover classification, as well as extend the methodology to other invasive species and ecosystems.