Assessment of Sentinel 2 derived phenometrics as predictor features for soil organic matter and pH mapping in a high-altitude Mediterranean forest

Soil mapping has been performed using different predictor features including climatic grids, terrain, or remotely sensed data. However, studies that consider dense time-series of remotely sensed imagery are scarce. Sentinel 2, the operational multispectral mission of European Space Agency, provides remotely sensed data with a 5-day revisit time and 10-m spatial resolution for visible and near infrared bands. Land Surface Phenology (LSP) and derived phenometrics could be obtained from Sentinel 2 time series. In complex forested ecosystems, these phenometrics could be useful predictor features for the spatial prediction of soil properties such as soil organic matter (SOM) or pH. The aim of this work was two folded, i) mapping SOM and pH with thirteen phenometrics derived from Sentinel 2 and terrain features using two machine learning algorithms (Random Forest (RF) and Support Vector Regression (SVR)) and two Feature Selection methods (Sequential Forward and Backward Selection) and ii) evaluating the contribution of LSP phenometrics for SOM and pH mapping through Feature Selection performance.

92 topsoil samples with SOM and pH data were collected in Sierra de las Nieves, southern Spain in 2019. The phenological features were extracted from a three-year time series of Enhanced Vegetation Index 2 (EVI2) computed from all available Sentinel 2 images of 30SUF tile for 2018-2020 period. The time series were smoothed using an asymmetrical gaussian method, and a 10% threshold-based method was used for phenometric extraction. Thirteen phenological features were extracted from the smoothed time series: amplitude, base value, end of season time, end of season value, large integral, left derivative, length of season, maximum value, middle of season, right derivative, small integral, start of season time and start of season value. Together with phenological data, elevation and twelve derived terrain features were used. The performance of two Machine Learning algorithms, Random Forest and Support Vector Regression, was evaluated within a framework with two Feature Selection methods, Sequential Forward Selection and Sequential Backward Selection.

The assessment of phenometrics for SOM and pH mapping highlighted the importance of middle of season for SOM, and Large Integral and End of Season value for pH prediction. Together with phenometrics, LS Factor for SOM and elevation and Channel Network Distance for pH were also found relevant. The performance of RF and SVR was similar for both soil properties, outperforming SVR in terms of R² for SOM modelling (SOM: R² of 0.06-0.20 and RMSE of 5.42-5.53; pH: R² of 0.20-0.37 and RMSE of 0.38-0.40). These results underpinned the suitability of Sentinel 2 time-series and LSP derived phenometrics for soil mapping in forested areas.