Geospatial Artificial Intelligence Detects Invasive Cactus Species across Large Heterogeneous Landscape of Laikipia, Kenya using Sentinel-2 Satellite Imagery

Abstract

Opuntia stricta, an invasive cactus species in Kenya's arid and semi-arid lands, has threatened grassland ecosystems and pastoralist livelihoods for decades. Remote sensing offers promise for detecting and monitoring these species, yet distinguishing them within mixed vegetation across expansive regions remains challenging. In this study, we employed Geospatial Artificial Intelligence (GeoAI) methods and Sentinel-2 satellite images to detect and map Opuntia stricta across Laikipia County's heterogeneous landscapes in Kenya. Our predictive modeling, grounded in a socio-ecological systems framework, involved extracting over 60 predictors encompassing biophysical and socioeconomic domains. These included time-series remote sensing indices (NDVI, EVI, GNDVI, NDWI, MSAVI2), topographic, climate, landscape structure, proximity variables, and regional socioeconomic indicators reflecting population and livestock density. Employing Recursive Feature Elimination (RFE), we identified the most effective feature subset and evaluated five machine learning algorithms: Random Forest, Support Vector Machine, XGBoost, Adaboost, and Multilayer Perceptron. Our best model achieved 90% accuracy in binary prediction (cactus vs non-cactus) and 85% in multi-class predictions (cactus, trees, shrublands, grasslands, bare land). Our analysis highlighted the significance of time-series indices capturing seasonal vegetation changes and specific spectral bands (green, red edge, red, and short-wave infrared) in distinguishing Opuntia stricta from other vegetation. These findings highlight the potential of GeoAI methods in accurately identifying individual invasive plant species across heterogeneous landscapes using moderate-resolution satellite imagery. Our results have implications for ecosystem management and restoration, offering guidance for the identification and monitoring of invasive plants, especially in remote and data-scarce regions.