Using resampled nSight-2 hyperspectral data and various machine learning classifiers for discriminating wetland plant species in a Ramsar Wetland site, South Africa

Abstract Mapping wetland ecosystems at the species level provides critical information for understanding the nutrient cycle, carbon sequestration, retention and purification of water, waste treatment and pollution control. However, wetland ecosystems are threatened by climate variability and change and anthropogenic activities; thus, their assessment and monitoring have become critical to inform proper management interventions. Contemporary studies show that satellite-based Earth observation (EO) has significant potential for achieving this task. While many multispectral EO data are freely and readily available, its broad spectral bands limit its utility in differentiating subtle differences among similar plant species. In contrast, hyperspectral data has a high spectral resolution, which is superior in discerning minute differences in similar plant species. However, this data is associated with high dimensionality and multicollinearity, which negatively affect the performance of traditional, parametric classification algorithms. To this end, machine algorithms are often preferred to classify hyperspectral data due to their robustness to various data distributions and noise. The current study compared the performance of three advanced machine learning classifiers, i.e., Support Vector Machine (SVM), Random Forest (RF), and Partial Least Squares Discriminant Analysis (PLS-DA), in discriminating four dominant wetland plant species, i.e., Crocosmia sp., Grasses, Agapanthus sp. and Cyperus sp. using simulated hyperspectral data from an upcoming sensor, i.e., nSight-2. The results revealed that SVM is superior, with an overall accuracy of 93.18% (and class-wise accuracies > 85%). In comparison, there were minor differences in the performances of RF and PLS-DA, i.e., 84.09% and 83.63%, respectively. Overall, the results demonstrated that all the evaluated classifiers could achieve acceptable mapping accuracies. However, SVM is more robust, providing exceptional accuracies, and should be considered for operational mapping once the sensor is in space.