Detecting the temporal trend of cultivated soil organic carbon content using visible near infrared spectroscopy

Monitoring changes in soil properties is essential to ensure ecosystem function and agricultural productivity. This study evaluated the ability of visible near infrared (Vis-NIR) spectroscopy to detect the temporal trend in soil organic carbon (SOC) content after 5 years in a 12 km2 agricultural catchment in western France. Partial least squares regression models were developed using soil samples from a local dataset collected in 2013 at two depths (198 samples at 0–15 cm and 196 samples at 15–25 cm) to predict SOC content of 111 new samples collected in 2018 at the same locations and at similar depths (0–15 cm and 15–25 cm). Two approaches, which differed in whether or not they considered the SOC content variability that can result from collecting soil samples at two depths, were applied. For both approaches, the potential benefit of “temporal spiking” was evaluated by adding 10% of 2018 samples to the 2013 dataset. The results showed that removing outliers and stratifying the calibration dataset by depth yielded the highest accuracy, with SOC RMSEP of 4.1 and 2.7 g.kg−1 for 0–15 and 15–25 cm, respectively. Moreover, temporal spiking improved five of eight predictions (stratifying or not the calibration dataset by depth, removing or not poorly predicted outliers), with increases in the ratio of performance to deviation of 0.10–0.44. Furthermore, comparing observed and predicted changes in SOC content showed that Vis-NIR spectroscopy estimated its trend over time in most cases.