Spatial modelling of Soil Organic Carbon fractions in a degraded coal-mining area through UAV and Sentinel-2

Traditional methods to acquire (geo)chemical data of Soil Organic Carbon (SOC) in soil rely on manual sampling, time-consuming and laborious chemical analyses, and subsequent mapping by geostatistical interpolation methods. In this study, we propose the use of UAV-RS and Sentinel-2 images, still partially supported by field sampling, for assessing and mapping different fractions of SOC using a regression study through Machine Learning (ML) techniques. This approach is exemplified in the postmining degraded soils of a vast former coal-mining area affected mainly by high degradation of Organic Matter (O.M).

Geochemical analyses by means of a TOC analyzer were conducted to monitor SOC fractions. Soil samples were dried and sieved through a 2-mm mesh to eliminate large particles. Two labile fractions of carbon (C_LAB) were obtained through cold-water extraction (C_CWE) and hot-water extraction (C_HWE); also, two removable carbon fractions (C_REM), humic and fulvic acids (C_HA and C_FA), were extracted; finally, the remaining recalcitrant organic carbon (C_REC) was measured in the residue of the previous extractions. TOC was estimated as the sum of C_LAB, C_REM and C_REC.

Spectral data were systematically recorded across a surface area covering 64 hectares within former open pits, involving natural, restored, and degraded zones. A UAV-RS P4-Multispectral platform, equipped with a camera featuring six individual sensors (RGB, blue, green, red, red-edge, and near-infrared), was used; five distinct bands between the visible and near-infrared spectra were obtained. Simultaneously, Sentinel-2 data were employed to acquire spectral information from satellite-borne sensors, thereby obtaining 12 single bands (aerosol, blue, green, red, 3 red edge, 2 NIR, water vapor, cirrus, and 2 SWIR). Given the limitations of information derived from individual bands, spectral indices—combinations of multiple bands through algebraic operations—were employed. Subsequently, two ML algorithms, specifically Random Forest (RF) and Partial Least Square (PLS), were applied to identify the most fitted model for each SOC fraction.

Results revealed that the utilization of non-parametric algorithms, specifically RF, yields a superior goodness of fit compared to parametric algorithms like PLS. The most favourable statistical outcomes were observed for fractions of non-labile organic carbon, with the optimal statistics achieved for C_REC, attaining an R² value of 0.70 and an RPD value of 1.83. When comparing data from UAV and Sentinel-2 sources, better results were found for UAV, this strongly suggesting that, in this study, spatial resolution holds greater relevance than spectral resolution.

This research was funded by the projects NATURESOIL (AEI/Spain, TED2021-130375B-I00) and Atlantic Risk Management Plan in Water and Soil (RiskAquaSoil 272-2016, Interreg Atlantic Area, EU).