Application of Multispectral Data and Machine Learning Technology in Estimating Soil Organic Carbon in Forest Ecosystems

In this study, combined with visible and near-infrared spectroscopy data and Fourier transform infrared spectroscopy data, machine learning methods were used to estimate soil organic carbon (SOC) of Eucalyptus plantations in southern China. Three machine learning models of support vector machine (SVM), random forest (RF), and deep neural network (DNN) in machine learning are selected, and their performance in SOC estimation is compared. The results show that the DNN model outperformed the SVM and RF models in accuracy, with the highest correlation with measured SOC values, with an R2 value of 0.95 and the lowest root mean square error and mean absolute error, respectively. Although the DNN model performs well in prediction accuracy, it has large computational resource requirements and poor model interpretability. The study also found that the estimation result of SOC was affected by the quality of spectral data and environmental factors. By optimizing the data collection and pretreatment process and exploring the integrated learning methods, this study provides a scientific basis and technical support for regional SOC estimation and carbon cycle research.