Skip to main content
ARS Home » Southeast Area » Fayetteville, Arkansas » Poultry Production and Product Safety Research » Research » Publications at this Location » Publication #431485
Research Project: Developing Best Management Practices for Poultry Litter to Improve Agronomic Value and Reduce Air, Soil and Water Pollution

Location: Poultry Production and Product Safety Research

Title: Free-access geospatial data and machine learning deliver high-resolution soil salinity maps in the Sonoran Basin and Range

Author
item VALENCIA-ORTIZ, MILTON - University Of Arkansas
item Ashworth, Amanda
item WINZLER, EDWIN - University Of Texas At Arlington
item Blackstock, Joshua
item MANCINI, MARCELO - Federal University - Brazil
item MUENICH, REBECCA - University Of Arkansas
item Owens, Phillip

Submitted to: ASABE Annual International Meeting
Publication Type: Abstract Only
Publication Acceptance Date: 1/15/2026
Publication Date: 2/2/2026
Citation: Valencia-Ortiz, M., Ashworth, A.J., Winzler, E., Blackstock, J.M., Mancini, M., Muenich, R., Owens, P.R. 2026. Free-access geospatial data and machine learning deliver high-resolution soil salinity maps in the Sonoran Basin and Range. ASABE Annual International Meeting, July 12-15, 2026. Indianapolis, Indiana.

Interpretive Summary:

Technical Abstract: The accumulation of salt on the soil surface can drastically affect crop productivity and the dynamics of natural ecosystems. These impacts are expected to intensify in the future, particularly in semi-arid to arid environments. The integration of machine learning (ML) with freely accessible geospatial variables can enable scalable extrapolation of ground-truth soil salinity data. In this study, 361 soil samples were collected across two Tribal Lands located in the Sonoran Basin and Range ecoregion, Arizona, USA. Then, electrical conductivity was measured using the saturated paste method. Next, 64 predictors were extracted from Sentinel-1, Sentinel-2, land surface temperature, apparent electrical conductivity, and supplemental geospatial data. Nearly 50 ML classifiers were tested to identify a single stable model. Recursive feature elimination was set to select 12 predictors, yielding an optimized model (OM). Inferential analysis was then applied to reduce them to six predictors, yielding a simplified model (SM). Cross-validation was performed on both OM and SM. Logistic regression showed stable accuracy across evaluated ML models. For OM and SM, comparable accuracy (0.68 ±0.04 and 0.68 ±0.05) was observed. Therefore, SM was used to make salinity predictions of 121,110 ha at 10-m resolution. According to the uncertainty analysis, the non-saline predicted class showed a larger area with uncertainties less than 0.2 than the saline class. The aridity index and Sentinel-2 features were the most important predictors. Findings demonstrate the model’s ability to extrapolate salinity data. This is essential for guiding appropriate agricultural development and advancing salinity mapping in arid regions.