Dataset and code for "Spatiotemporal monitoring of cropland soil organic carbon changes from space"

Dataset for:

Spatiotemporal monitoring of cropland soil organic carbon changes from space (10.1111/gcb.17608)

Tom Broeg, Axel Don, Martin Wiesmeier, Thomas Scholten, Stefan Erasmi

Contact: tom.broeg@thuenen.de

This repository contains the datasets and scripts to run the spatiotemporal SOC model, described in “Spatiotemporal monitoring of cropland soil organic carbon changes from space”. It can be used to reproduce the main findings of the paper, including the cross-validation of the SOC model, the validation of the predicted SOC trends, as well as the prediction of the SOC trend map from 1986 to 2021.

1. Data files

(1) SOC reference data: The file SOC_data.csv contains information on the SOC measurements, conducted in the research area between 1986 and 2022.

“SOC”: Measured SOC content (0 -15 cm) in g kg^-1

“Sample_Year”: The year in which the respective SOC samples were collected

“Dataset”: Soil sampling programs; LTM = Long-term Soil Monitoring Program, HDB = Humus Database, BZE = German Soil Inventory (Bodenzustandserhebung Landwirtschaft)

“Field_ID”: Identifier for each LTM, HDB, and BZE plot, used to combine the SOC data and model features

(2) Model features: The file training_data.csv contains the dynamic and static features from all sampling programs (LTM, HDB, and BZE) that were used to train the spatiotemporal SOC model. The file LTM_data.csv contains the model features for all SRC pixels intersecting the LTM plots (~30x30m) and was used to predict the SOC trends.

“RED”, ”GREEN”, ”BLUE”, “NIR”, “SWIR1”, “SWIR2”: Spectral bands from the Landsat/Sentinel 2 soil reflectance composite (SRC)

“BCC”, “GCC”, “RCC”: Blue, Green, and Red Chromatic Coordinates, derived from the SRC bands (multiplied by 1000)

“NBR2”: Normalized Burn Ratio 2, derived from the SRC bands (multiplied by 1000)

“srtm_30_crop”: Elevation data from the Shuttle Radar Topography Mission (SRTM)

“BGL”: Soil landscape units (“Bodengroßlandschaften”) from the German Federal Institute for Geosciences and Natural Resources

“Field_ID”: Identifier for each LTM plot, used to combine the SOC data and model features

“Sample_Year”: Sampling year in which the dynamic SRC was generated

“Sample_Nr”: Identifier for the SRC pixels within the LTM plots: 1 = Center; 2-5 = Corners

(3) Model validation: The files cv_training_data.csv and cv_LTM_data.csv contain the cross-validated SOC predictions that were used to evaluate the spatiotemporal model and predicted SOC trends.

“SOC_Predicted”: Cross-validated SOC predictions in g kg^-1

“Sample_Year”: Sampling year of the respective SOC predictions

“Field_ID”: Identifier for each LTM, HDB, and BZE plot

“Sample_Nr”: Identifier for the SRC pixels within the LTM plots: 1 = Center; 2-5 = Corners

(4) Raster data: The files SRC_1986.tif – SRC_2021.tif contain the dynamics soil reflectance composites that were generated every five years from 1986 to 2021. The additional static covariates are provided in static.tif. The file SOC_prediction_raster.tif contains eight layers with the corresponding SOC maps from 1986 to 2021. The generated SOC trend map, based on these predictions, is provided in SOC_trends_raster.tif.

“Layers SRC_1986.tif – SRC_2021.tif”: See Model features (2)

“Layers SOC_prediction_raster.tif”: SOC predictions from 1986 to 2021 in g kg^-1

“Layer SOC_trends_raster.tif”: Derived SOC trend from 1986 to 2021 (Pearson r, multiplied by 1000)

2. Code

(1) Model training and validation: The R code SOC_model.Rmd contains details on the implementation and validation of the spatiotemporal SOC model. The SOC reference data (SOC_data.csv) is combined with the model features for each sampling year and Field_ID (training_data.csv) to conduct a support vector machine (SVM) and 10-fold cross-validation. The cross-validation is repeated for all SRC pixels intersecting the LTM plots (LTM_data.csv) to derive and validate the SOC trends, based on the long-term SOC measurements.

(2)  Prediction of the SOC trend map: The R code SOC_map.Rmd contains details on the production of the SOC maps, based on the spatiotemporal SOC model and the dynamic SRC, located in SRC_1986.tif – SRC_2021.tif. The individual SOC maps from 1986 to 2021 (SOC_prediction_raster.tif) are combined the derive the final SOC trend map for the research area (SOC_trends_raster.tif)