Core Concepts
The proposed knowledge-embedded contrastive learning (KnowCL) framework unifies supervised, unsupervised, and semi-supervised hyperspectral image classification into a scalable end-to-end framework, leveraging labeled and unlabeled data to achieve superior performance compared to existing methods.
Abstract
The paper presents a novel knowledge-embedded contrastive learning (KnowCL) framework for hyperspectral image (HSI) classification. The key highlights are:
Data Processing Pipeline:
Adopts a disjoint data sampling strategy to realistically partition training and test sets, avoiding inflated evaluation metrics.
Applies various data augmentation techniques to generate diverse labeled and unlabeled samples.
Unified Framework Design:
Encompasses supervised, unsupervised, and semi-supervised learning approaches within a single scalable framework.
Utilizes a backbone network (e.g., ViT, ResNet) with supervised and contrastive heads to extract features from labeled and unlabeled data.
Introduces a novel loss function that adaptively fuses supervised and contrastive losses to leverage both labeled and unlabeled samples.
Experimental Evaluation:
Outperforms state-of-the-art supervised, unsupervised, and semi-supervised methods on benchmark HSI datasets.
Demonstrates the effectiveness of the proposed framework in terms of classification accuracy, model complexity, and training efficiency.
Provides detailed analysis on the impact of hyperparameters like crop size, batch size, and number of bands.
The KnowCL framework presents a unified and scalable solution for HSI classification, effectively exploiting labeled and unlabeled data to achieve superior performance compared to existing methods.
Stats
The number of training samples for the DFC2018 dataset is only 7% of the total labeled samples.
The number of training samples for the UP dataset is 30% of the total labeled samples.
The number of training samples for the Salinas dataset is 20% of the total labeled samples.
The number of training samples for the Dioni dataset is 50% of the total labeled samples.
Quotes
"The proposed framework based on this pipeline is compatible with all kinds of backbones and can fully exploit labeled and unlabeled samples with expected training time."
"We develop a new semi-supervised learning paradigm for HSI classification, which is very useful in practical applications. This framework is achieved by a new loss function based on multi-task learning, designed to adaptively combine supervised and contrastive losses."