Abstract: Organic carbon, as one of the important components of soil, is of great significance in assessing soil quality and stability. In karst areas, understanding the distribution characteristics of soil organic carbon content can identify potential soil erosion risk areas and provide a scientific basis for optimizing land use and formulating effective soil and water conservation measures. We collected limestone soil samples under different land use types in Guinan karst, captured soil color images by using five smartphones to extract soil color parameters, and used a spectrometer to obtain soil spectral information. We combined machine learning methods and linear algorithms i.e., artificial neural network (BPNN), support vector machine (SVM), and random forest (RF), as well as the linear algorithm partial least squares regression (PLSR), to establish an inverse prediction model for organic carbon content. We used the coefficient of determination, root mean square error, and relative analytical error as the model accuracy evaluation indices, to screen and specify the smartphones and the corresponding prediction models applicable to soil organic carbon content in the region and the prediction models under the spectrometer method. The results showed that the modeling results of five smartphones based on the four modeling methods presented different effects: Redmi Note11T pro+>IQOO Neo7 SE>Huawei nova 5Z>realme X7 pro>iPhone X. The integrated modeling effects of the multispectral data collected by the five smartphones and the hyperspectral data collected by the spectrometer were consistent. The SVM accuracy assessment coefficient was the best and the modeling effect showed superiority, followed by BPNN, RF and PLSR, compared with PLSR, the machine learning algorithm showed a better prediction effect. Combined with the scatter plot of model estimation, the model predictions were more dispersed when soil organic carbon content was lower than 10 g·kg^-1 and more concentrated when the soil organic carbon content was higher than 10 g·kg^-1. This study could provide theoretical support and practical basis for understan-ding the spatial distribution characteristics of soil organic carbon content in karst area, which would be a basis for solving the problem of soil erosion and improving the agricultural production environment in this area.

Key words: karst, soil organic carbon, soil color, spectral inversion, machine learning