Dual-modal fusion of hierarchical image features and spectral data for efficient quantitative analysis of mineral elements in rice (Oryza sativa L.) leaves.

Abstract

Rapid and accurate quantification of mineral elements in plants facilitates the optimization of cultivation strategies and provides theoretical support for heavy metal pollution control. Compared to traditional chemical detection methods, laser-induced breakdown spectroscopy (LIBS) offers rapid, simultaneous multi-element analysis. However, the quantitative accuracy of LIBS is often hindered by challenges such as sample heterogeneity and the inherent matrix effects arising from the physical and chemical properties of samples. These limitations highlight the need for innovative approaches to improve the reliability and precision of LIBS-based elemental quantification. In this study, we proposed a low-cost image-spectroscopy dual-modal rapid detection system combined with a dual-modal hierarchical fusion network (DMH-FNet). Compared with a standalone LIBS system, the quantification performance improved for the seven elements, namely P, Ca, Mg, Zn, Mn, K, and Si. During the validation phase, feature map visualization was used to interpret the feature extraction process of DMH-FNet. The results indicate that the model shifted its focus from low-level features of ablation crater details to high-level global features of the sample. Subsequently, SHapley additive exPlanations (SHAP) was used to explain the decision-making process of the optimal quantitative model and visualize key image features. The results demonstrate that DMH-FNet efficiently extracts features highly correlated with ablation crater information using its neural network capabilities and enhances the quantification of mineral elements through complementary fusion with LIBS spectral features. This study is the first to leverage the superior feature extraction capability of neural networks to capture valuable information from ablation images, thereby improving the quantification performance for multiple mineral elements. In conclusion, the proposed detection system, with its low-cost equipment, real-time data acquisition, and DMH-FNet, enables simultaneous, rapid, and accurate prediction of multiple mineral element contents.

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.