Spectral-Spatial Ensemble Learning for Invasive Robinia Pseudoacacia Detection Using UAV-Based Hyperspectral Imaging

Abstract. The National Heritage Administration of Korea designates and manages particular non-native species, as well as highly reproductive native plants with strong environmental adaptability, as invasive plants to preserve the unique natural landscapes within cultural heritage sites. However, investigating and managing large-scale cultural heritage areas—such as palaces and fortresses—requires considerable time and labor, and these efforts are further hindered by challenging terrain. This study investigates the use of hyperspectral imaging (HSI), acquired via unmanned aerial vehicles (UAVs), as an efficient approach for monitoring the distribution of black locust (R. pseudoacacia), a representative invasive species. HSI provides rich spectral information by continuously measuring reflectance across a wide range of wavelength bands. We utilize HSI data comprising 150 spectral bands to detect the presence of black locust in the Gongsanseong and Busosanseong fortress areas. To address the limitations of benchmark-based models, such as poor generalizability and overfitting to dataset-specific features, we propose an ensemble approach that integrates the strengths of multiple learning models. This includes neural networks designed to capture both spectral and spatial features, allowing for complementary processing of complex spectral patterns and spatial contextual information. From a numerical study, the proposed method achieves robust detection performance for target species, even in heterogeneous environments.