No Detector Training Needed: Lens Finds Look-Alike Targets Across Satellite Tiles From a Single Click
Synopsis
For a satellite mosaic spanning thousands of tiles, an analyst previously had to collect examples and train a detector before asking where a given kind of target appears. Lens instead encodes each tile with a frozen self-supervised vision transformer, applies two linear corrections that remove positional bias and the area-wide common component, and matches the averaged features around a click against a byte-quantized index by cosine similarity. On the DFC2020 test set, a single click reaches 0.09 to 0.80 pixel F1 across eight land-cover classes, mean 0.39, and beats a pairwise INSID3 baseline on six of them at roughly four hundred times lower query cost.
Interpretation
Semantic search over a satellite mosaic can run with no training at all: the analyst clicks one example and the system returns every similar location in the area, with the class defined by that click. The prior route turned each question into a supervised learning task, collecting examples, annotating, training and tuning a detector over weeks, and starting over for the next target class. On the 5,128-tile DFC2020 test set, a single click reaches 0.09 to 0.80 pixel F1 across eight land-cover classes with a mean of 0.39; against a pairwise INSID3 baseline it wins six of the eight classes, and the baseline is additionally given each image pair its own best threshold.
After encoding each tile into dense descriptors, the system projects them onto the subspace orthogonal to a positional basis and then subtracts the mean descriptor of the whole area, so similarity measures departure from the area average. Dense features carry a content-independent positional component, so patches at the same grid coordinates match spuriously across tiles; and when an area is uniform, descriptors cluster so tightly that cosine similarity loses its range. The positional basis is estimated once from a constant image and cached; over 300 DFC2020 tile pairs, a patch's best cross-tile match falls at the same grid position 5.2% of the time before debiasing and 0.8% after, close to the 0.39% chance rate.
Indexing and caching bring interactive queries to milliseconds: centered patch features go into an 8-bit quantized Faiss index that supports range search over the whole mosaic. A pairwise scan of the corpus costs about 188 seconds per query, which cannot support dragging a threshold slider. A range search over 1.31 million patches takes 436 ms; on five Sentinel-2 sites, a click returns a rendered overlay in 6 to 10 ms, of which the search itself is 1.7 to 2.7 ms.
Perspective
The demonstrated scope is per-class accuracy on the DFC2020 test set and interactive operation on five Sentinel-2 sites spanning agricultural, industrial and desert terrain; all timings come from single-thread CPU execution on one machine. Accuracy is measured against land-cover products rather than instance annotations, so for object classes the scores are lower bounds.
Centering is not always beneficial: above a target prevalence of roughly 13% it costs accuracy, and nothing in the pipeline detects that condition. Similarity scores are relative to the indexed area, so the same value in two areas does not mean the same thing; compositing takes a median over the date window, collapsing time into one image, and only three of Sentinel-2's thirteen bands are used.
