计算机科学
人工智能
高光谱成像
域适应
领域(数学分析)
模式识别(心理学)
适应(眼睛)
遥感
特征提取
计算机视觉
特征学习
自然语言处理
训练集
作者
Yuhao Xie,Ronghua Shang,Jinhong Ren,Jingyu Zhong,Jie Feng,Songhua Xu
标识
DOI:10.1109/tgrs.2026.3674804
摘要
Hyperspectral domain adaptation (HSDA) recontextualizes unsupervised domain adaptation for the spectrally rich and structurally complex task of hyperspectral image classification. However, most existing HSDA methods confine source supervision to an auxiliary role for stabilizing training, overlooking its potential to act as a central driving force in the adaptation process. Moreover, global alignment at the domain level has not been effectively emphasized in existing methods. To address these limitations, we propose a source-driven global-local prototype contrastive learning (SGLPCL) framework that fully leverages source supervision as a central driver of adaptation through hierarchical prototype alignment, fostering both domain-invariant and semantically coherent representations. To this end, SGLPCL constructs robust node representations from spectral-spatial-semantic embeddings and learns global representations through a multi-head self-attention graph network. It then refines semantic pseudo-labels at both global and local levels using multi-scale distance metrics, leveraging source supervision for reliable labels. Finally, a global-local prototype contrastive learning framework is established, with an adaptive temperature scaling matrix, utilizing source supervision to calibrate prototype similarities and enhance contrastive learning robustness and discriminability. Comprehensive evaluations across three representative hyperspectral benchmarks substantiate the superiority of SGLPCL, which attains remarkable improvements in accuracy and Kappa coefficient over ten leading methods.
科研通智能强力驱动
Strongly Powered by AbleSci AI