Predictive soil mapping using historic bare soil composite imagery and legacy soil survey data

• Published in: Geoderma 2021-11, Vol.401, p.115316, Article 115316
• Main Authors: Sorenson, P.T., Shirtliffe, S.J., Bedard-Haughn, A.K.
• Format: Article
• Language: English
• Subjects: Bare Soil Composite; Predictive Soil Mapping; Remote Sensing; Soil Organic Carbon
• ISSN: 0016-7061; 1872-6259
• DOI: 10.1016/j.geoderma.2021.115316

•Models could be built using a bare soil composite image from Landsat 5.•Soil organic carbon model was driven by near infrared and visible light bands.•Clay model was primarily dependent on the shortwave infrared bands.•Cation exchange capacity model was driven by shortwave and near infrared bands. There is an increasing need for detailed soil property maps to support land use planning, soil carbon accounting, and precision agriculture. While soil maps exist in Saskatchewan, they are at coarse scales (1:100,000), which are not always suitable for detailed soil management. One emerging technique for predictive soil mapping is the use of bare soil composite imagery derived from multi-temporal satellite imagery. This study focused on using bare soil composite imagery along with legacy soil data (1958–1998) with high location uncertainty to predict soil organic carbon, clay, and cation exchange capacity. The bare soil composite images were created from Landsat 5 imagery (1985 to 1995) using Google Earth Engine. Predictive models were built using a Random Forest model for each parameter and evaluated using a 75–25 train-test split. The soil organic carbon model had an R2 value of 0.55 with a root mean square error (RMSE) of 0.67 percent, with the near infrared and visible light bands being the most important features in the model. The clay predictive model has an R2 of 0.44 and a RMSE of 5.0 percent, with the shortwave infrared bands being most important. The cation exchange capacity model had an R2 of 0.50 with a RMSE of 5.7 meq 100 g−1, with the shortwave and near infrared bands as the most important predictors. Based on these results, bare soil composite imagery represents a valuable covariate for predictive soil mapping in the Canadian Prairies. This work also illustrates that for regions with extensive adoption of conservation farming practices, satellite imagery should be obtained for time periods before these practices were adopted from the months of the year where crop residues have decomposed. By combining historical soil survey data with historical imagery, maps of legacy soil properties can be generated to make comparisons against with modern data for applications such as monitoring soil organic carbon change over time.