Evaluating UAV spectroscopy for monitoring soil organic carbon in agricultural fields

Recent literature shows a strong increase in publications using UAV technology [1], and more specifically for soil-related applications such as the prediction of SOC on agricultural land. This technology holds great promises for SOC mapping, particularly in the context of carbon farming, for which temporal monitoring at field-plot scale is required. However, before this technology can be widely used, key questions remain regarding accuracy, temporal sensitivity, and cost-effectiveness compared to satellite remote sensing or geostatistical approaches.

Using a two-season field monitoring over an 11.25 ha-plot, with three replicates of five different tillage practices and a total of 75 sampling points, we aim to address these questions. During the first campaign (October 2024), a multispectral UAV (10 bands from 444 to 842 nm, VNIR) was used, while during the second campaign (May 2025), a hyperspectral UAV (~500 bands from 400 to 2500 nm, VNIR-SWIR) was deployed. To our knowledge, this is the first attempt to map SOC using VNIR-SWIR hyperspectral UAV. In parallel, for both seasons, soil samples were collected for laboratory SOC analysis and spectral measurements under controlled conditions (dried and sieved samples).

We used machine learning models (PLSR, RF, SVM) to predict SOC, comparing spectra derived from UAV imagery, Sentinel-2 (S2) data, and laboratory spectra. For this purpose, the dataset was split into 2/3 for calibration and 1/3 for validation, and this procedure was repeated randomly 100 times. The same data partitioning was used to evaluate a kriging approach.

The first result shows that surface SOC concentration is strongly dependant on tillage practices, with a mean seasonal change of 1.2 g.kg⁻¹ (±1.3 g.kg⁻¹) over a 10-20 g.kg⁻¹ range. This also raises questions about the importance of acquiring spectral data close in time to soil sampling. For UAV, S2, and kriging approaches, model performance was lower in May than in October, with decreases of 1.1, 0.8, and 0.5 in RPIQ, respectively. This suggests more favorable surface and/or sky conditions in October, despite wetter soils and sparse vegetation.

In October, multispectral UAV achieved high prediction accuracy comparable to laboratory spectroscopy (RPIQ ≈ 3.1), followed by S2 and kriging (RPIQ = 2.7 and 2.3, respectively). The same ranking was observed in May, however, the performance of the hyperspectral UAV decreased substantially and became similar to S2 (RPIQ = 2.1 vs. 1.9). Adverse weather conditions may partly explain this decline. Reducing calibration sampling density did not significantly degrade UAV accuracy, indicating potential for cost reduction. Our results suggest that future UAV-based studies should systematically compare their results with alternative methods and not only report prediction performance, but also explicitly address cost-effectiveness and temporal monitoring constraints.

[1] S. A. H. Mohsan, N. Q. H. Othman, Y. Li, M. H. Alsharif, and M. A. Khan, “Unmanned aerial vehicles (UAVs): practical aspects, applications, open challenges, security issues, and future trends,” Intell. Serv. Robot., vol. 16, no. 1, pp. 109–137, Mar. 2023, doi: 10.1007/s11370-022-00452-4.