Mapping within-field soil health variations using apparent electrical conductivity, topography, and machine learning

Authors: Adhikari, Kabindra; Smith, Douglas; Collins, Harold; Hajda, Chad; ACHARYA, BHARAT - State Of Oklahoma; Owens, Phillip

Submitted to: Agronomy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 4/21/2022
Publication Date: 4/24/2022
Citation: Adhikari, K., Smith, D.R., Collins, H.P., Hajda, C.B., Acharya, B.S., Owens, P.R. 2022. Mapping within-field soil health variations using apparent electrical conductivity, topography, and machine learning. Agronomy. 12. Article 1019. https://doi.org/10.3390/agronomy12051019.
DOI: https://doi.org/10.3390/agronomy12051019

Interpretive Summary: Field map of soil health status could help farmers in making soil and crop management decisions. This study applied machine learning technique to map within-field variability of soil health indicators, and soil health index based on Haney Soil Health Tool using apparent electrical conductivity (ECa), and topographic information. Results showed that ECa and topography were correlated with soil health measurements. There was a strong correlation between soil health index and ECa and that the later could be fairly estimated using ECa data collected on-the-go using DualEM sensor.

Technical Abstract: Assessment and mapping of within-field soil health variation could help farmers and managers to finetune farm resources for profit maximization. To address these issues, a study was conducted in Texas Blackland soils with the following main objectives: i) to assess and map within-field variability in soil health using machine learning; iii) to evaluate the usefulness of local topography and apparent electrical conductivity (ECa) as predictors of soil health; and (iii) to quantify the relationship between ECa and soil health index, and use ECa collected on-the-go to estimate SHI distribution. Two hundred and eighteen topsoil (0-10 cm depth) samples were collected in a 35-m grid, and a soil health index (SHI) based on Haney Soil Health Tool was determined. Apparent electrical conductivity data were collected on-the-go with a DualEM sensor, and local topographic information from a digital elevation model. A machine learning model using random forest (RF) algorithms was applied to predict and map one day CO2, organic carbon, organic nitrogen, and SHI where ECa, and terrain attributes representing local topography were used as predictors. Furthermore, empirical relationship between SHI and ECa was established and the correlation between them was mapped across the field. Results showed that the study area was variable in terms of one day CO2, organic carbon, organic nitrogen, SHI, and ECa distribution. The ECa, and terrain attributes such as wetness index, multiresolution valley bottom flatness and topographic position index were among the top predictors of one day CO2, organic carbon, organic nitrogen, and SHI. The prediction model was found robust to predict one day CO2, organic carbon, organic nitrogen (R2 between 0.24-0.90), and SHI (R2 between 0.47-0.90). Overall, we observed a moderate to strong spatial dependency of one day CO2, organic carbon, organic nitrogen, and SHI in the study area, and these factors could impact yield variability at within-field scale. The study confirmed the applicability of easy to obtain ECa as a good predictor of SHI, and the predicted maps at high resolution (5 m x 5m grid) which could be useful in SSCM management decisions within these types of soils.