. . . . . . . . . . . . "2026-04-18"^^ . . "Within-domain few-shot learning is recommended over cross-domain transfer for satellite imagery classification" . "Applicable to optical satellite imagery (Sentinel-2, Landsat) for land cover and habitat classification tasks where labeled examples are scarce. Tested on 10 broad EuroSAT land cover categories at 10 m ground resolution using RGB bands only." . "Not tested on fine-grained habitat discrimination such as distinguishing vegetation subtypes within Natura 2000 sites. Only RGB bands used — near-infrared and shortwave infrared bands, which are critical for vegetation analysis, were not included. Results may differ for non-optical sensors (SAR, LiDAR) or very high resolution imagery. The within-domain experiment used a different base/novel class split than a real monitoring scenario would require." . "For Earth observation researchers with limited labeled satellite data: if any labeled satellite imagery is available for your region (even for different classes than your target), use within-domain few-shot learning — train on common classes and classify rare ones. If no satellite training data exists at all, use an off-the-shelf ImageNet-pretrained model as a feature extractor for initial screening, then invest in labeling a small satellite dataset to improve accuracy. Complex meta-learning pipelines are not necessary — standard supervised pretraining achieves comparable results." . "Four experiments comparing within-domain and cross-domain few-shot learning on Sentinel-2 satellite imagery show that within-domain transfer (training on common satellite land cover classes, classifying rare ones) achieves 82% accuracy with 5 labeled examples per class, while cross-domain transfer from everyday photographs achieves 67–76% depending on backbone architecture and training method. The domain gap between photographs and satellite imagery reduces accuracy by 6–15 percentage points. Supervised pretraining on everyday photographs matches episodic meta-learning with 12 times less training time, but both cross-domain approaches remain below within-domain accuracy." . . "few-shot learning - machine learning approach that enables a system to learn new tasks or recognize new objects from only a few examples or demonstrations, rather than requiring extensive data" . "remote sensing - acquisition of information about an object or phenomenon without making physical contact with the object, especially the Earth" . "land cover - nature of the physical material at the surface of the earth" . "transfer learning - research problem in machine learning (ML) that focuses on storing knowledge gained while solving one problem and applying it to a different but related problem" . "Anne Fouilloux" . "2026-04-18T18:55:18.984Z"^^ . . . . . . . . . "RSA" . "MIGfMA0GCSqGSIb3DQEBAQUAA4GNADCBiQKBgQDWv2pJnmDsBOq8OlT1aSvYXSuWT34WOp4FYqEzdnn2F0kqzcFevBqWGZDxJWC0lqCrDEuNfp2QFyPe/+nES9dlHGYIhqPi68fwK6ZiNUotRFxXou+rjFznVvUxtCL8Ede79EBHwWN61QtwSIcU12bLoZsNPFlqQASQ93BJuKlihwIDAQAB" . "FTGbwGUbqi+nCLOqOznThPR9MOLOcTesabQLyyR5desYZTqLy/St4qAbY50Y781/RcoQGvufwkJEp73g0bXf1UKhSykRwfC/NrC01/kh5emK/uuwxgbVd4+avSqwx3cXSa0i5Esx2Xv3J4WFDm8qXw2GhtmWcdDXkthZHQ6MiW4=" . . .