Improved soil property prediction models for rice paddies using measured nutrients

Abstract Traditional laboratory analyses of soil texture and soil organic matter (SOM) are time consuming and labor‐intensive, so they are impractical to be measured at the same scale as routine nutrient testing. The objective of this study was to determine soil clay, sand, and SOM across a large number of samples by improving mixed‐land‐use predictive models proposed by Drescher et al. in 2024, creating a rice ( Oryza sativa L.)‐specific set of models that integrates soil pH, Mehlich‐3 extractable nutrients, and estimated cation exchange capacity (EstCEC). With a clustering analysis, a soil dataset containing 179 samples from major rice producing states in the United States was split into a training set (80%) for rice model development and a testing set (20%) for validation. Another dataset of 111 samples from Arkansas was used to compare the performance of rice models with mixed‐land‐use models. After validation, a high‐accuracy clay model ( R 2 = 0.84; RMSE = 68.14 g kg −1 ) was obtained using pH, phosphorous (P), potassium (K), calcium (Ca), and magnesium (Mg). The sand model containing pH, Ca, Mg, and EstCEC fit with moderate accuracy ( R 2 = 0.36; RMSE = 89.94 g kg −1 ). The best SOM model relied on pH, P, K, Mg, and EstCEC ( R 2 = 0.80; RMSE = 4.28 g kg −1 ). The Arkansas rice soil dataset showed that rice models enhanced SOM prediction ( R 2 increased from 0.78 to 0.81), and they improved the overall soil textural classification accuracy to 70% versus 58% by mixed‐land‐use models. While our models are suitable for clay and silt‐dominated classes, such as clay, silt loam, clay loam, and silty clay loam, they may not be for several sand‐dominated classes. This study provides a tool by which to efficiently and inexpensively estimate key soil physicochemical properties for agricultural decision‐making but the overall utility of our models for rice soil textural classification is limited at present.