|
Title
|
|
|
|
A contrastive learning method for multi-label predictors on hyperspectral images
| |
|
Author
|
|
|
|
| |
|
Abstract
|
|
|
|
| Self-supervised contrastive learning is increasingly acknowledged as an effective approach to mitigate the challenges posed by limited annotated data. We introduce a two-stage methodology that extends current approaches, targeting the downstream task of multi-label classification in hyperspectral remote-sensing imagery. In the initial stage, we employ a contrastive learning approach to train a base encoder and a projection neural network, thereby learning data patterns without relying on annotations. The effectiveness of the encoder is bolstered as it is guided by a contrastive loss function to maximize the similarity between the generated embeddings. In the second stage, we harness the power of the pre-trained encoder to channel its hidden representations into a multi-label classifier. Our empirical validation demonstrate that this method surpasses fully supervised alternatives. The observed improvements are attributed to the strategy of training the encoder alongside the classifier, thereby refining its adaptability to the feature space of the classifier. |
| |
|
Language
|
|
|
|
English
| |
|
Source (book)
|
|
|
|
2023 13th Workshop on Hyperspectral Imaging and Signal Processing : Evolution in Remote Sensing (WHISPERS), 31 October 2023 - 02 November, 2023, Athens, Greece
| |
|
Publication
|
|
|
|
IEEE
,
2023
| |
|
ISBN
|
|
|
|
979-83-503-9557-0
| |
|
DOI
|
|
|
|
10.1109/WHISPERS61460.2023.10430726
| |
|
Volume/pages
|
|
|
|
p. 1-5
| |
|
Full text (Publisher's DOI)
|
|
|
|
| |
|