Probing the lighting sensitivity of image encoders with repeat drone imagery: A case study of plant height estimation

Image encoders provide a strong backbone for tasks such as retrieval, classification, and depth estimation, and recent releases tailored to remote sensing, such as DINOv3 with SAT pretraining [Siméoni et al., 2025], promise improved performance on ecologically important applications. It is uncertain whether such encoders yield robust features under variable illumination, where shadows and hue shifts can obscure relevant plant structure. To address this, we developed a drone imagery dataset with high-resolution RGB captures of the same site at three time points in a single day, paired with plant canopy height models. Using these data we identified the subspace of embeddings dominated by lighting variation and progressively projected embeddings away from lighting subspace components. Across both DINOv2 and DINOv3, canopy height prediction remained stable until more than 80\% of the lighting variance was removed, after which performance degraded sharply, with a pronounced error spike when the full lighting subspace was eliminated. These results suggest that while much of the lighting variance is nuisance, the final fraction contains useful textural and chromatic cues. DINOv3-SAT consistently outperformed the general-purpose DINOv2, maintaining ~1 cm lower error until complete removal of the lighting subspace. We release the [$\text{Drone-LSR}$](https://huggingface.co/datasets/mpg-ranch/drone-lsr) dataset on Hugging Face under a Creative Commons 4.0 license to facilitate further exploration of lighting sensitivity in image encoders for remote sensing.