A semi-supervised domain adaptation framework with attention-enhanced ResNet50 for LIBS-based soil classification under spectral distribution shift.

BACKGROUND: Laser-induced breakdown spectroscopy (LIBS), as a rapid and non-destructive analytical technique, has been widely applied in soil classification research. In recent years, the integration of LIBS with deep learning has significantly improved the performance of soil classification tasks. However, spectral distribution shifts resulting from variations in soil physical and chemical properties pose a major challenge to model performance. Therefore, there is an urgent need for an effective method to address the spectral distribution shifts and mitigate the impact on the analysis results. RESULTS: This study proposes a semi-supervised domain adaptation (SSDA) framework for cross-domain classification of LIBS spectra. Specifically, the one-dimensional spectral data after feature dimensionality reduction are encoded into two-dimensional images using the Gramian Angular Field (GAF) technique to enhance feature representation. Then, a deep convolutional neural network based on SE-ResNet50 is designed to focus on informative features through attention mechanisms. On this basis, a SSDA framework is developed by integrating model-based transfer learning (MBTL) and pseudo-label learning (PLL) to gradually improve the model's adaptability to the target domain. In this framework, MBTL is used to initialize the target model using knowledge from the source domain, while PLL expands the supervision signals by leveraging high-confidence predictions. Experimental results demonstrate that the proposed method achieves a classification accuracy of 97.67% on the target domain test set, significantly outperforming comparative models. SIGNIFICANCE AND NOVELTY: This study validates the effectiveness of the proposed method in mitigating spectral distribution shifts and successfully addresses the spectral distribution shift problem caused by differences between soil samples. Based on this method, the performance of LIBS technology in cross-domain soil classification has been significantly improved, demonstrating its application potential in the field of soil analysis.