Machine learning-based high-resolution soil mapping for precision crop stress management

Overview

This solution leverages machine learning and geospatial modeling to generate high-resolution, accurate maps of essential soil properties needed for within-field precision management of crop stress. By combining high-dimensional, multi-scale digital surface derivatives with extensive soils training data, the approach produces predictive soil maps that are significantly more accurate and far finer in resolution than currently available large-extent soil databases such as USDA-NRCS SSURGO. The technology addresses a critical need in precision agriculture: detailed soil information at the sub-field scale to manage abiotic crop stress effectively.

Technical specifications

• Resolution advantage: Maps produced at 3-meter resolution (each cell covering 9 square meters), compared to SSURGO's mean delineation size of approximately 101,170 square meters.
• Soil properties mapped: Soil organic matter, topsoil thickness, and soil texture (sand, silt, and clay content).
• Modeling approach: Cubist algorithm trained on high-dimensional, multi-scale digital surface derivatives, with multi-temporal agricultural aerial imagery, satellite imagery, and PRISM climate data as additional predictors to capture vegetative and soil response variables.
• Validation results: Independent validation in Iowa demonstrated reductions in error for silt content (from 9.5% to 7.1%) and sand content (from 12.3% to 7.8%), with both models averaging an R-squared of 0.66. Carbon predictions for 0 to 60 cm depth showed errors of 0.5% to 0.9% and R-squared values of 0.32 to 0.47. Topsoil thickness prediction error was nearly halved (from 25.7 cm to 13.7 cm) with an R-squared of 0.80.
• Future depth coverage: Maps will be created at six depth increments down to 2 meters, with associated error maps included.

Technology readiness level

The technology has been validated against the current standard for large-extent soil maps, demonstrating significant accuracy improvements over the USDA-NRCS SSURGO database. Future validation plans include expanding sampling to 180 sites across central Iowa (adding 60 new sites to the existing 120) to a depth of 2 meters, with sites strategically selected based on preliminary error distributions to improve model performance. Accuracy will be assessed via 10-fold cross-validation and a 5% independent validation set. The approach is at an advanced stage of development with demonstrated validation results, ready for further refinement and potential deployment in precision agriculture applications.