Cross-Domain Heterogeneous Hyperspectral Image Classification Based on Meta-Learning with Task-Adaptive Loss Function
- Resource Type
- Conference
- Authors
- Jin, Yuheng; Ye, Minchao; Xiong, Fengchao; Qian, Yuntao
- Source
- 2023 13th Workshop on Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS) Hyperspectral Imaging and Signal Processing: Evolution in Remote Sensing (WHISPERS), 2023 13th Workshop on. :1-5 Oct, 2023
- Subject
- Computing and Processing
Signal Processing and Analysis
Metalearning
Adaptation models
Signal processing algorithms
Classification algorithms
Task analysis
Optimization
Hyperspectral imaging
Hyperspectral image
cross-domain classification
meta-learning
- Language
- ISSN
- 2158-6276
The scarcity of labeled samples poses a significant challenge in hyperspectral image (HSI) classification. Cross-domain classification offers a potential solution by leveraging information from a source domain with abundant labeled samples to classify the target domain, even when only limited labeled samples are available. Our work is to classify the target domain where there are classes that are not in the source domain, which is a difficult task. This paper presents a method called cross-domain meta-learning with task-adaptive loss function (CD-MTALF) for heterogeneous HSI classification. CD-MTALF is a transfer learning-based method specifically designed to address the issue of limited labeled samples in the target domain. CD-MTALF adopts a meta-learning approach based on optimization. It consists of two optimization tasks: 1) Inner-loop optimization: adapting the base learner to a new task using a learning algorithm. 2) Outer-loop optimization: training the base learner, meta-learner, and meta-learnable loss function and employing an adaptive weighting strategy to balance the losses between the source and target domains. CD-MTALF aims to enhance the model’s generalization performance by incorporating both inner-loop and outer-loop optimizations. Experiments on publicly available HSI datasets demonstrate the efficacy of CD-MTALF in improving classification performance.