Open AccessSelf‐supervised learning with randomised layers for remote sensing
Author(s) -
Jung Heechul,
Jeon Taegyun
Publication year - 2021
Publication title -
electronics letters
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.375
H-Index - 146
eISSN - 1350-911X
pISSN - 0013-5194
DOI - 10.1049/ell2.12108
Subject(s) - computer science , random forest , binary classification , binary number , artificial intelligence , entropy (arrow of time) , principle of maximum entropy , supervised learning , cross entropy , semi supervised learning , machine learning , logistic regression , layer (electronics) , task (project management) , pattern recognition (psychology) , remote sensing , mathematics , support vector machine , artificial neural network , engineering , chemistry , physics , arithmetic , organic chemistry , systems engineering , quantum mechanics , geology
This letter presents a new self‐supervised learning approach based on randomised layers for remote sensing. Our method is basically based on the Tile2Vec approach, which is one of the state‐of‐the‐art self‐supervised learning approaches for remote sensing. Unlike the original Tile2Vec algorithm, we reformulate the triplet loss as a classification loss. We use several fully connected layers with binary cross‐entropy loss instead of no fully connected layers with triplet loss of the original Tile2Vec. We observe that not updating the fully connected layers is more helpful in obtaining more robust representations. The proposed algorithm is verified and evaluated by applying it to a cropland data layer classification task. The experimental results show that our approach is superior to the original Tile2Vec approach in all experiments based on random forest, logistic regression, and multi‐layer classifiers.