Λεπτομέρειες βιβλιογραφικής εγγραφής
| Τίτλος: |
Few-shot soil provenance analysis by laser-induced breakdown spectroscopy combining intra-class interpolation-enhanced prototypical learning. |
| Συγγραφείς: |
He, Yihua1,2 (AUTHOR), Ding, Yu1,3 (AUTHOR) dingyu@nuist.edu.cn, Pang, Maoyuan1,3 (AUTHOR), Wang, Yuhang1,3 (AUTHOR), Li, Jinyi1,3 (AUTHOR), Yu, Weiye1,3 (AUTHOR), Li, Xiangchu1,3 (AUTHOR), Xu, Jianan1,3 (AUTHOR), Tan, Qiang1,3 (AUTHOR), Liu, Xinxin1,3 (AUTHOR), Li, Luning4 (AUTHOR) liluning@mail.sitp.ac.cn |
| Πηγή: |
JAAS (Journal of Analytical Atomic Spectrometry). Aug2026, Vol. 41 Issue 8, p2935-2947. 13p. |
| Θεματικοί όροι: |
*Laser-induced breakdown spectroscopy, *Soil classification, *Feature extraction, *Machine learning, *Dimensional reduction algorithms, *Data augmentation |
| Περίληψη: |
Rapid and reliable soil provenance analysis is important for agriculture, environmental monitoring, and forensic traceability, but LIBS-based soil classification remains challenging when labeled calibration samples are limited. An uncertainty-weighted prototypical network with intra-class interpolation enhancement was developed for few-shot LIBS soil classification. LIBS spectra from 14 soil origins were first compressed from 8190 wavelength variables to 500 representative spectral regions by feature agglomeration, reducing spectral redundancy while retaining discriminative information. To improve prototype robustness under sparse calibration conditions, synthetic intra-class features were generated by linear interpolation between samples from the same class to enrich the local class distribution in the embedding space, while sample contributions were adaptively weighted according to their similarity to class prototypes to reduce the influence of uncertain or ambiguous samples. Under the most challenging condition, with only 10% of spectra used for training, feature agglomeration improved the ProtoNet accuracy from 66.98% using full spectra to 81.27%, and intra-class interpolation further increased the accuracy to 85.71%. Across training-set proportions of 10%, 20%, 30%, and 40%, the final model achieved accuracies ranging from 85.71% to 93.57% and consistently outperformed SVM, RF, and MLP classifiers. These results demonstrate that spectral feature compression, intra-class augmentation, and uncertainty-aware metric learning can improve the accuracy of LIBS soil identification with limited calibration samples, providing a sample-efficient analytical strategy for soil provenance analysis. [ABSTRACT FROM AUTHOR] |
| Βάση Δεδομένων: |
Academic Search Index |