نسخة أولية وصول مفتوح
T-JEPA: A Temporal Joint-Embedding Predictive Architecture for Learning Better Remote Sensing Representations
Earth observation (EO) data provide rich temporal supervision, yet existing remote sensing foundation models mainly exploit sequential observations through imposing predefined pairwise relations or aggregating holistic reconstruction context. We seek to further exploit the sparse and nonuniform temporal sampling inhere …