Testing the efficacy of hyperspectral (AVIRIS-NG), multispectral (Sentinel-2) and radar (Sentinel-1) remote sensing images to detect native and invasive non-native trees.
Invasive alien species threaten tropical grasslands and native biodiversity across the globe, including in the natural mosaic of native grasslands and forests in the Shola Sky Islands of the Western Ghats. Here, grasslands have been lost to exotic tree invasion (Acacias, Eucalyptus, and Pines) since the 1950s, but differing invasion intensities between these species and intermixing with native species constitutes a major challenge for remotely sensed assessments. In this study, we assess the accuracy of three satellite and airborne remote sensing sensors (Sentinel-1 radar data, Sentinel-2 multispectral data and AVIRIS-NG hyperspectral data) and three machine learning classification algorithms to identify the spatial extent of native habitats and invasive tree species. We used the support vector machine (SVM), classification and regression trees (CART), and random forest (RF) algorithms implemented on the Google Earth Engine platform. Results indicate that AVIRIS-NG data in combination with SVM produced the highest classification accuracy (98.7%). Fused Sentinel-1 and Sentinel-2 produce 91% accuracy, while Sentinel-2 alone yielded 91% accuracy; but only with higher coverage of ground control points. The hyperspectral data (AVIRIS-NG) was the only sensor that permitted distinguishing recent invasions (young trees) with high precision. We suspect that large areas will have to be mapped and assessed in the coming years by conservation managers, NGOs to plan restoration or to assess the success of restoration activities, for which a choice of sensors may have to be made based on the age of invasion being mapped, and the quantum of ground control data available.