Workshop: Scale to See: Scalable AI for Scientific Imaging and Spatiotemporal Data
Authors: Junlin Guo, Sharmin Majumder, Isaac Lyngaas, John Lagergren, and Xiao Wang (Oak Ridge National Laboratory)
Abstract: High-resolution plant imaging enables detailed characterization of plant morphology, but dense scientific imaging analysis remains limited by costly pixel-level annotations, large image dimensions, and variable imaging conditions. This work proposes a scalable in-domain self-supervised foundation model for high-resolution poplar imagery. A masked autoencoder with a ViT backbone is pretrained on more than 10 million multi-species plant image tiles using distributed training. The learned representations are benchmarked for fine-grained dense prediction and coarse global retrieval under limited supervision and realistic downstream conditions. Across limited annotations, cross-view variation, and resolution degradation, the in-domain FM achieves a Mean Dice of 0.8686 and a Pooled Dice of 0.8959, outperforming an ImageNet-pretrained MAE by 0.0694 and 0.0613, respectively. Increasing pretraining scale further improves dense feature transfer, suggesting that scalable in-domain pretraining can reduce domain gaps in high-pixel-dimension scientific imaging.
Back to Scale to See: Scalable AI for Scientific Imaging and Spatiotemporal Data Archive Listing
Back to Full Workshop Archive Listing