Can a model trained on satellite data really find brambles on the ground? (toao.com)

Researchers led by Gabriel Mahler tested whether a model trained on satellite-derived embeddings can locate brambles on the ground. Using TESSERA earth-representation embeddings (via the geotessera library) paired with iNaturalist labels, the team built a simple ensemble (logistic regression + k‑nearest neighbors) to predict bramble hotspots. They then did rapid field validation around Cambridge—stopping at model high-confidence locations near Milton Community Centre, Milton Country Park, residential plots and a nature reserve—and consistently found substantial bramble where the model predicted hotspots. The takeaway for the AI/ML community is twofold: satellite-learned embeddings from Sentinel-1/2 can transfer to fine-scale vegetation detection when features are large and unobscured, enabling lightweight classifiers to perform well in the field; conversely, partial canopy cover or small understorey brambles are harder to detect, reflecting limits of optical/radar-derived embeddings. The team collected geo-tagged photos for further validation and noted practical prospects for on-device, human-in-the-loop active learning (mobile phones feeding new labeled points back to simple models), while future improvements could come from more validation data, higher-resolution sensors (or LiDAR), and incorporating occlusion-aware training.