Prediction of Soil Organic Carbon Content in Spartina alterniflora by Using UAV Multispectral and LiDAR Data

Abstract

Soil organic carbon (SOC) is an essential component for plant growth and a pivotal factor in the global carbon cycle. Spartina alterniflora (S. alterniflora), an invasive species characterized by high primary productivity and rapid carbon sequestration capabilities, exerts a substantial impact on SOC concentrations. The precise quantification of SOC content in S. alterniflora is extremely importance. Based on 73 measured samples, along with multispectral imagery and LiDAR data collected via unmanned aerial vehicles, machine learning techniques, including random forest (RF), support vector machine (SVM), and extreme gradient boosting (XGBoost), were employed to predict the SOC content of S. alterniflora and map its spatial distribution. We compared the predictive performance of these different machine learning algorithms to identify the most effective one. The results show that the following. 1) The prediction accuracy is improved by classifying the data into three types: unlodging S. alterniflora (ULSA), lodging S. alterniflora (LSA), and mudflats. 2) XGBoost outperformed RF and SVM in accurately predicting SOC content, with R²; values of 0.743 for ULSA, 0.731 for LSA, and 0.705 for mudflats; 3) In the XGBoost models constructed for ULSA, LSA, and mudflats, spectral features contributed 75.7%, 73.1%, and 63.1%, respectively, with the normalized difference vegetation index emerging as the most critical spectral feature. Slope aspect (AS) was identified as the most influential topographic feature. 4) The spatial distribution of SOC exhibited marked heterogeneity, with higher SOC content in ULSA and lower in mudflats, demonstrating a gradient of decreasing SOC content from land to sea. These results hold significant implications for the study of SOC content in S. alterniflora.