The domain gap challenges artificial intelligence models in Earth Observation, degrading performance when patterns shift between the data encountered during training and deployment. Although the existence of the gap is acknowledged, predicting this performance degradation is challenging, especially for tasks such as ship segmentation in satellite imagery, which are inherently imbalanced. This thesis demonstrates that traditional domain gap metrics are inadequate for this task due to their inability to handle class imbalance.

To overcome this, a new metric is proposed: the positive difference of confidences. By focusing only on the model's confidence in the positive (ship) class, it ignores insignificant background changes. Tested across 22 source-target domain combinations, the metric​ proved a powerful predictor of target performance, with a correlation of 0.78 and a mean absolute error of 0.06, significantly outperforming existing metrics. The positive difference of confidences​ offers an accurate method for ensuring performance of ship segmentation models.