Self-supervised satellite imagery representations encode physically meaningful environmental signals (ERA5 variables) that correlate with downstream task performance, particularly for agriculture and disaster domains.
Understanding contrastive representation learning through alignment and uniformity on the hypersphere
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
fields
cs.CV 2years
2026 2representative citing papers
Supervised and contrastive pretraining yield stronger linear separability than masked reconstruction or self-distillation on a three-class emerald grading task, with reconstruction improving under nonlinear probes.
citing papers explorer
-
Probing Geospatial SSL Representations with Environmental Signals
Self-supervised satellite imagery representations encode physically meaningful environmental signals (ERA5 variables) that correlate with downstream task performance, particularly for agriculture and disaster domains.
-
Pretraining Objective Matters in Extreme Low-Data FGVC: A Backbone-Controlled Study
Supervised and contrastive pretraining yield stronger linear separability than masked reconstruction or self-distillation on a three-class emerald grading task, with reconstruction improving under nonlinear probes.