Industry Perspective: EarthDaily

Scientific-Grade Earth Observation: From Constellation to Foundation Models — Data Quality as the Missing Variable

Diego Domingos Della Justina Data Scientist/BR Team Manager

Remote sensing foundation models are increasingly constrained not by architecture, but by the data they are trained on. Most available datasets trade off at least one of three critical dimensions: temporal density, spatial resolution, or spectral depth. Using a high-quality proxy dataset sharing the core characteristics of the EarthDaily Constellation — near-daily revisit, ~5m resolution, multi-spectral coverage — we examine how each dimension affects downstream model performance. In a corn harvest date forecasting task, reducing acquisition frequency by 50% (to approximate Sentinel-2 revisit) increases median error from 7.1 to 11.9 days (+67%). Training on 5m native imagery outperforms bicubic-upsampled 10m data by 24%. We also show that foundation model pretraining on high-quality data reduces annotation dependency: with 10,000 labels, a pretrained model outperforms training from scratch. In a complementary multi-sensor fusion experiment, expanding inputs from RGB to the full sensor suite yields a 2x IoU improvement on crop classification, suggesting that spectral breadth carries substantial signal currently absent from most training pipelines. We discuss Territory Insights — EarthDaily's operational crop intelligence product — as an applied example of what becomes tractable when all three data dimensions are addressed simultaneously, spanning automated crop stage tracking, disease risk indexing, and harvest forecasting at field scale. Finally, we introduce the EarthDaily Constellation (daily global coverage, 5m, 22 bands including SWIR and thermal, CEOS-ARD compliant), currently launching, and outline research collaboration opportunities through the Pioneer Program.