Why do classification models go wrong? The importance of adaptations and acclimations in driving landscape-level spectral variation in Fremont cottonwood

Spatially explicit predictions of species distributions can inform ecosystem processes and conservation, particularly under global change. While imaging spectroscopy could enable accurate species classifications, accuracy generally declines outside training regions, limiting its utility for regional-scale mapping. To investigate mechanisms constraining classification generalizability (e.g., spatial autocorrelation, local adaptation), we used National Ecological Observatory Network Airborne Observation Platform imaging spectroscopy data collected across riparian systems in Arizona, Colorado, and Utah. We extracted canopy spectral data of Populus fremontii (Fremont cottonwood), a foundation riparian tree known to form locally adapted ecotypes across its range, and spatially co-occurring species from seventeen 6 × 6 km sites. Combining this library with site-level environmental data, and support vector machine (SVM) models, we observed that environmental, not geographic, distance between training and test sites limited classification generalizability. Specifically, differences in mean annual temperature, winter precipitation, and spring precipitation, key drivers of local adaptation of P. fremontii, were associated with lower classification accuracy (∼50% lower). We then evaluated specific wavelength regions for improved generalizability. Classification models using only near-infrared (750–1400 nm) and shortwave infrared (1400–2500 nm) outperformed those using full-spectrum models in regions not represented in the training data, consistent with lower heritability in visible and red-edge wavelengths. In conclusion, spatially structured spectral phenotypes of P. fremontii, shaped by local adaptation and acclimation to environmental conditions, reduced species classification generalizability. By integrating ecology into remote sensing workflows, such as spectral band selection, we can improve species classification accuracy, thereby advancing scalable biodiversity monitoring and conservation efforts.