PhenoEmbed is a self-supervised model for learning 256-dimensional seasonal representations of individual tree crowns from multispectral UAV time series. It uses contrastive learning and masked reconstruction on HeideBench, an 18-date benchmark from Dölauer Heide. On 5,885 crop-safe crowns, the first two principal components explain 25.1% of embedding variance, while nearest-neighbor retrieval reaches a median top-1 cosine similarity of 0.946. The reported ablations indicate that contrastive loss, masked reconstruction, and explicit seasonal time features each influence the learned embedding geometry. The work positions the embeddings as reusable representations, but downstream gains under seasonal shifts remain to be tested.
No heat snapshots are available in the last 24 hours.