Spatiotemporal Machine Learning
Predictions of soil organic carbon density and soil pH are based on using spatiotemporal Machine Learning. For more details about how models are fitted and used to generate predictions, please refer to:
Hengl, T., Sorenson, P., Parente, L., Cornish, K., Battigelli, J., Bonannella, C., … & Nichols, K. (2023). Assessment of soil organic carbon stocks in Alberta using 2-scale sampling and 3D predictive soil mapping. FACETS, 8, 1-17. https://doi.org/10.1139/facets-2023-0040
Parente, L., Sloat, L., Mesquita, V., Consoli, D., Stanimirova, R., Hengl, T., ... & Stolle, F. (2024?). Mapping global grassland dynamics 2000—2022 at 30m spatial resolution using spatiotemporal Machine Learning, submitted to Scientific Data. https://doi.org/10.21203/rs.3.rs-4514820/v2
Predictions of soil carbon, soil pH, and soil types are based on using a large collection of combined training points. The image below indicates the distribution of training points used (currently only Conterminous USA is covered):
Some previous predictions of soil types and soil properties (static models) at 100-meter spatial resolution are available from:
Ramcharan, A., Hengl, T., Nauman, T., Brungard, C., Waltman, S., Wills, S., & Thompson, J. (2018). Soil property and class maps of the conterminous United States at 100‐meter spatial resolution. Soil Science Society of America Journal, 82(1), 186-201. https://doi.org/10.2136/sssaj2017.04.0122
The main sources of training points include:
- USDA-NRCS Kellogg Soil Survey Laboratory (KSSL) soil database as imported into OSSL: 31,809 samples
- USGS NGS points: 9,141 samples
- SOils DAta Harmonization database (SoDaH) points: 5,677 samples
- Potash et al. (2023): 3,977 samples
Relationship between soil carbon content and soil depth.
- Image: A simple pedo-transfer function was used to fill-in gaps for soil organic carbon density (kg/m3) where soil carbon content (g/kg) + soil depth are used to estimate soil carbon density for all soils with <0.5% of SOC (low values). Relationship between soil carbon density and soil carbon content is otherwise more complex for soils with >1% of soil carbon.
The processing and integration of the soil samples used to build models are explained in detail in:
Safanelli, J. L., Hengl, T., Parente, L. L., Minarik, R., Bloom, D. E., Todd-Brown, K., ... & Sanderman, J. (2024?). Open Soil Spectral Library (OSSL): Building reproducible soil calibration models through open development and community engagement. Submitted to PLOS One. https://doi.org/10.1101/2023.12.16.572011
A subset of this data is available as open data and can be used to validate models and predictions (see Kaggle.com competition “ESA EO4SoilProtection 2024: Predicting SOC Density”).