Grunnleggende konsepter
衛星画像における自己監督学習方法の重要性と効果を示す。
Sammendrag
衛星画像のラベル付きデータが限られているため、自己監督アルゴリズムが有用であることが示唆されています。S3-TSSは、時間次元で自然な拡張を活用する新しい自己監督学習技術を提案しています。この手法は、現在の最先端の手法よりも優れた結果を示しました。さまざまな実験も行われ、その成果が比較されました。また、他の研究や手法との関連性も明らかにされています。
Statistikk
Satellite images have a higher temporal frequency.
S3-TSS method outperformed baseline SeCo in four downstream datasets.
SeCo dataset contains 100,000 images with 5 seasonal variants each.
EuroSAT dataset consists of 27,000 labeled images into ten classes.
AID dataset includes 10,000 images distributed into 30 aerial scene types.
UCMerced Land Use Dataset has 21 classes with 100 images each.
WHU-RS19 Dataset provides high-resolution remote sensing images up to 50cm with 19 classes.
Sitater
"Self-supervised learning, a new paradigm, has drawn a lot of interest since it takes advantage of algorithms’ intrinsic capacity to produce supervisory signals from unlabeled data."
"Satellite images undergo natural transformations over time, including stationary alterations such as lightning, solar radiation, weather conditions, and day-night transitions."
"Our goal was to learn a student model that was better than the teacher model and hence more numbers and difficult augmentations were given as input to the student model."