Integration of Machine Learning and Remote-Sensing Reflectance Spectra for Regional SOC Mapping: A Case Study on the Agricultural Lands of the Loess Plateau

Soil organic carbon (SOC) is a key component of the terrestrial carbon cycle and is essential for soil fertility, directly influencing climate change and human well-being. However, it remains unclear how different spectral data sources and machine learning (ML) models can jointly influence SOC prediction performance, especially across large and heterogeneous agricultural landscapes. This study systematically evaluates combinations of spectral data sources, feature-selection methods, and ML models for SOC prediction in the farmland of the Loess Plateau (LP), a region characterized by fragmented croplands and limited carbon stock data. Based on large-scale field sampling of topsoil (0–5 cm) and measurements of SOC content and reflectance spectra from 460 samples, we evaluated 52 combinations of ML models and three categories of input data, including hyperspectral data, resampled multispectral data, and multispectral data combined with environmental variables. The results show that the hyperspectral-based model achieved the highest accuracy ( R 2 = 0.97 in validation), but its reliance on regional-scale hyperspectral datasets restricts its practical applications. In contrast, the multispectral-environmental integration model delivered scalable performance, offering a feasible pathway for regional SOC mapping. The mapping results in this case study show that the SOC content in the farmland of the LP ranges from 0.002 to 22.26 g·kg –1 , with low SOC levels predominantly distributed in the loess hilly gully region and the central parts of the sandy and agricultural irrigation region. This study establishes a systematic framework for SOC mapping in heterogeneous agricultural landscapes and offers practical support for carbon management in agroecosystems.