Near-infrared spectroscopy for predicting soil properties and fertilizer recommendations in Korean cropping systems

This is the authors' abstract. We don't add key points for this paper.

Accurate soil analysis plays a crucial role in fertilizer recommendations. Conventional soil analysis is time-consuming, requires chemicals, and does not provide real-time soil nutrient status. In this context, spectroscopy offers a promising alternative, enabling simultaneous assessment of multiple soil properties. Therefore, the study objective was to evaluate the potential of Near-Infrared (NIR) spectroscopy to predict soil properties and its applicability to fertilizer recommendations for seven crops. A total of 1167 soil samples were collected from different land-uses. The samples were analyzed by conventional method, and NIR and prediction models were developed using five algorithms: Random Forest (RF), Cubist, Partial Least Squares Regression (PLSR), Support Vector Machine (SVM), and Memory-Based Learning (MBL). The MBL outperformed other models in predicting soil attributes. The results showed strong prediction accuracy for total nitrogen (TN) (R 2 = 0.84, RMSE = 0.05, RPD = 2.50, Bias = 0.00), organic matter (OM) (R 2 = 0.81, RMSE = 9.07, RPD = 2.27, Bias = 0.20), and electrical conductivity (EC) (R 2 = 0.79, RMSE = 2.05, RPD = 2.27, Bias = 0.12), while moderate for pH, available phosphorus, potassium (K), magnesium (Mg), and calcium (Ca) (R 2 < 0.68, RMSE < 310.58, RPD < 1.74, Bias < 0.92). Compared with wet-chemistry results, average differences in fertilizer recommendations for seven crops based on NIR were within acceptable limits (N: 3.6%, P: 7.7%, and K: 9.7%). Overall, this study demonstrates the potential of NIR as a cost-effective and reliable tool for soil assessment, enabling efficient fertilizer recommendations and supporting sustainable agriculture.