Breaking01 / AI
Satellite image vectors predict event completion dates

Researchers reported in a preprint published on 13 Nov that their satellite event prediction model achieved a median completion-date error of 3 days (n=undisclosed, predictions before completion), using differences between image vectors; the paper is not peer-reviewed, and its abstract reports no independent replication results.
The paper, Predicting the Needle in a Petabyte Scale Haystack: Open-Vocabulary Event Anticipation in Satellite Imagery, describes using a vision-language model, which processes images and text, to represent images from different times as vectors: numerical representations of their content. The researchers report using differences between those vectors to identify changes, collect events without manual labelling and build a dataset for training a prediction model.
The model's performance in retrieving predicted events was 3.6 times that of baseline models (n=undisclosed, baseline comparison reported in the abstract), according to the paper.
The researchers argue that earlier systems cannot jointly predict where a change occurs, what it will become and when it will finish without restricting predictions to predefined event types. Their approach links event collection to prediction-model training by searching for changes in meaning between images through vector differences.
The abstract does not disclose the evaluation sample size or prediction errors by event type.