Keeping up with a growing satellite archive without wrecking the old ordering — adapters plus ranking distillation for continual retrieval

1Learn something new and the old correspondence breaks

Earth-observation satellites keep taking pictures, and the archive keeps growing. A system that lets you search images with text has to keep learning as new vocabulary and new features appear in it. But learning more breaks what was aligned before. This is a familiar property of machine learning: the more new material is absorbed, the more old material fades. In retrieval the breakage has a particular shape.

It is not that one record becomes unfindable; it is that the ordering of many candidates, the structure of what counts as closer, gets distorted.

2Where to intervene

1. 1The archive growsNew features and new phrasings appear
2. 2Learn moreLeft alone, the space aligning images with text distorts
3. 3Add adaptersOn the visual side, integrate coarse regions with fine detail and anchor updates to the original alignment space
4. 4Separate the expertsOn the text side, split shared meaning from specialized parts to restrain global drift
5. 5Distill the rankingKeep the earlier model frozen as a teacher so the previous ordering carries over

What stands out in this design is that the original model is never overwritten. Small components are added on both the visual and the text side, and the effect of updating is confined to them. The earlier model is then kept frozen as a teacher, so the past carries forward in the form of an ordering. The device for not forgetting sits on the side that preserves old weights, not on the side that learns new ones.

3How often does adding show up

Work about taking an existing model and turning it toward another setting, rather than training from scratch, accounts for nearly a fifth of the AI papers held here. This study belongs to the part of that group which both adds something and preserves something. The requirement to keep up with data that keeps arriving is not confined to Earth observation; it applies wherever records accumulate.

4Do not misread what makes it hard

In remote sensing the same feature looks very different at different resolutions and under different capture conditions, and the distribution itself moves over time. The authors describe this as worsening the distortion of the alignment space. Read the other way, continual learning is hard here for reasons beyond the general problem of forgetting. Unless the difficulties are separated, the remedy misses.