Waititu, Julius MainaMundia, Charles NdegwaSichang, Arthur W.2023-09-142023-09-142023Waititu, J. M., Mundia, C. N., & Sichangi, A. W. (2023). Observation and Geoinformation. International Journal of Applied Earth Observation and Geoinformation, 119, 103307.https://doi.org/10.1016/j.jag.2023.103307http://ir-library.ku.ac.ke/handle/123456789/26952ArticleLantana Camara L. (LC) invasive species has not been successfully mapped due to inadequate spectral information. This study aimed at assessing the performance of leaf-level in-situ hyperspectral data and derived indices in discriminating LC among co-occurring species during the dry and wet seasons. In addition, the performance of simulated Sentinel-2 bands, Sentinel-2 derived indices and machine learning algorithms in discriminating it was explored. Spectrally distinct features for species discrimination were selected using the guided regularized random forest (GRRF) and their separability quantified with Jeffries–Matusita distance method. We found that ratio-based and difference indices constructed with first and second-order derivative hyperspectral reflectance wavelengths perfectly separated LC from co-occurring species in the dry and wet seasons with ≥ 97% of separability accuracy. Similarly, a set of derived ratio-based and difference Sentinel-2 indices yielded > 95% and < 80% of LC separability accuracy in wet and dry seasons respectively. The SVM with radial basis function algorithm fitted with selected continuum removed derivative reflectance (band depth) narrow-bands yielded the highest overall accuracy (OA) of 84% and a Kappa of 0.75 for the dry season while the same algorithm fitted with selected first derivative narrow-bands yielded an OA of 82% and Kappa of 0.66 for the wet season. Conversely, the regularized logistic regression yielded the highest performance (OA of 77% and Kappa of 0.62) when fitted with combined selected Sentinel-2 variables for the dry season while the gradient boosting machine (GBM) fitted with combined Sentinel-2 variables had the highest performance (OA of 75% and Kappa of 0.51) for the wet season. These findings have important implications on the upscaling of LC’s derived leaf-level indices to canopylevel and subsequent LC classification with hyperspectral and Sentinel-2 imagery datasets over heterogeneous environments.enLantana camaraInvasive speciesSpecies discriminationHyperspectralSentinel-2Machine learning algorithmsArticle