Evaluating the Effectiveness of Sentinel-2 Multispectral Remote Sensing Data in Identifying Weeping Willow (Salix babylonica) in a Heterogeneous Urban Environment

Abstract

Weeping willow (Salix babylonica) trees are alien invasive deciduous plants originally introduced to South Africa for decorating landscapes. Over time, their dense canopies have been associated with shading of understory vegetation, potentially affecting local biodiversity and water-dependent ecosystems. Their occurrence along urban riverbanks and other water-dependent environments, together with their potential effects on native vegetation and water-dependent ecosystems, makes their identification and spatial monitoring particularly relevant. However, the ability of medium-resolution Sentinel-2 imagery to distinguish weeping willows from spectrally similar urban vegetation remains poorly understood, particularly in heterogeneous urban environments such as Johannesburg. This study evaluated the capability of Sentinel-2 multispectral imagery and machine learning algorithms in identifying willow trees in the heterogeneous urban environment of Johannesburg, South Africa. It was hypothesized that Sentinel-2 spectral bands and derived indices could distinguish willow trees from co-occurring vegetation, but that spectral similarity and mixed pixels would constrain classification performance. Two hundred weeping willow observations from selected locations based on prior knowledge of the study area were recorded and supplemented by a few verified records from the Global Biodiversity Information Facility (GBIF). Ten Sentinel-2 Level 2A spectral bands corresponding to the sampling period were included: natural color bands, Red-edge, Near-Infrared, and Shortwave Infrared bands, with seven additional derived spectral indices, including EVI, NDRE, MNDWI, MCARI, Clred-edge 1 and 2, and PRI, to enhance vegetation, chlorophyll, moisture, and physiological information. Spectral values were extracted at ground-reference locations, and the data were divided into 60% training and 40% validation samples. Random Forest (RF) and Gradient Boosted Classification (GBC) models were developed and evaluated using accuracy metrics, variable importance, logistic GLM validation, spatial autocorrelation, and ROC analysis. RF and GBC achieved moderate overall accuracies of 75% and 74%, respectively, with kappa coefficients of 0.71 and 0.70 for identifying willow trees. Furthermore, both models achieved a producer’s accuracy of 60% and user’s accuracy of 54% for weeping willows, indicating moderate discrimination and reflecting the difficulty of separating willows from other vegetation with similar spectral characteristics. The moderate performance was likely influenced by the limited reference-sample representation and the medium spatial resolution of Sentinel-2, where individual pixels may contain mixtures of tree canopy and surrounding vegetation or non-vegetated surfaces, as well as spectral overlap among vegetation classes with similar chlorophyll, moisture, and canopy characteristics. EVI, PRI, and MNDWI were among the most influential predictors, emphasizing the importance of biophysical and physiological variables in invasive species detection.