Crop-field scale prediction of soil organic carbon in a complex terrain and its implications for nitrogen recommendation

Abstract

This study aimed to evaluate the potential of spectral data in the Vis–NIR–SWIR region, combined with terrain and soil variables, for predicting soil organic carbon (SOC) using different machine learning algorithms, and to analyze the impact of these estimates on nitrogen recommendations for maize ( Zea mays L.). A total of 179 soil samples were collected in a 21 ha agricultural field, with 144 used for model calibration and 35 for independent validation. SOC was determined by dry combustion and adopted as the reference method. Models were developed using Partial Least Squares Regression (PLSR), Random Forest (RF), and Cubist (CUB) algorithms. Spectral data were used alone or integrated with topographic covariates derived from spatial resolutions of 5, 10, and 30 m. PLSR with spectral smoothing, integrating spectral and topographic covariates at 10 m resolution, showed the best performance (Concordance Correlation Coefficient, CCC = 0.95; Root Mean Square Error, RMSE = 3.30 g kg⁻¹). The topographic covariates with the highest relative importance were slope position, LS factor, valley depth, elevation, and channel network distance. Spectral bands accounted for more than 90% of the relative importance in explaining SOC variability. Spectrally predicted SOC values showed high agreement with the reference method and resulted in statistically equivalent nitrogen recommendations, with reduced fertilization costs compared to the conventional method. This study demonstrates that Vis–NIR–SWIR spectroscopy is a reliable and operational tool for SOC estimation and nitrogen management at the field scale in agricultural landscapes with high topographic variability.