Generalizing EO Spectral Representations to Unseen Sensors: A Soil Spectroscopy Study

Earth-observation models increasingly aim to operate across sensors with different spectral configurations, yet the ability to accept arbitrary band layouts does not necessarily imply that the learned representation will generalize to a sensor absent during training. We study this question in a controlled setting using dense soil VNIR-SWIR spectra resampled to Sentinel-2, Landsat-9, Landsat-Next, and EnMAP, allowing spectral-configuration shift to be isolated from atmospheric, spatial, and temporal differences. A wavelength-aware Transformer is evaluated under a strict leave-one-sensor-out protocol, where the target sensor is excluded from both pretraining and supervised regression. We augment pretraining with randomly sampled generic spectral configurations to reduce specialization to the observed sensors. Across the four held-out sensors, mean soil organic carbon R^2 improves from -0.13 to 0.53, while clay R^2 improves from 0.31 to 0.66, relative to fixed-configuration pretraining. These results show that sensor-flexible architectures alone do not guarantee unseen-sensor generalization and that increasing spectral-configuration diversity during pretraining substantially improves transfer.