Abstract:

A model trained in one geographic regime and deployed in another is a routine situation in applied machine learning, and the usual question asked about it is whether transfer works. The practical question is different: given that transfer is imperfect, how many labelled examples from the new regime are needed before local training is the better option, and what exactly do the source data buy while labels are scarce?

We answer both for a frozen gradient-boosted detector of mineral-dust events trained on 900 hyperspectral satellite acquisitions over North Africa and deployed over two regimes it never saw: the Middle East and East Asia. Three strategies are compared at seven label budgets under granule-disjoint validation with all arms scored on identical test folds: the frozen model, a model trained on local data only, and a model trained on source and local data together.