Leveraging Multimodal and Large Language Models: Advances and Prospects for Rapid Prediction of Soil Organic Carbon

The rapid and accurate prediction of soil organic carbon (SOC) content is of great significance for the study of carbon cycling and the evaluation of soil fertility. Nowadays, with the development of spectroscopy and intelligent technologies, related technologies have been widely used in the prediction of soil elements due to their advantages of low cost, non-destructiveness, real-time performance, and rapid determination. In this article, the advantages of visible-near-infrared (VNIR) and hyperspectral imaging (HSI) spectral technologies are first described. Feature extraction and multi-feature fusion methods based on spectral data are analyzed, the role of multi-feature fusion in improving prediction accuracy is emphasized, and the establishment of efficient and reliable prediction models is discussed. Subsequently, multi-feature fusion methods for spectral data are explored, and it is pointed out that single feature extraction methods have limitations, requiring the combination of deep learning and manual feature extraction methods. In terms of modeling methods, the foundation of SOC prediction is introduced, especially the research on the prediction of soil profile organic carbon based on VNIR and HSI, as well as the application of multimodal models in agriculture. The important role of the combination of multimodal models and large language models in improving the level of intelligent agricultural production is highlighted. Therefore, the application of multimodal models in agriculture is summarized. LLMs may serve as knowledge-integration and decision-support interfaces that help interpret SOC prediction outputs, integrate agronomic knowledge, and convey information to the end users.