Vis–NIR spectroscopy and machine learning methods to diagnose chemical properties in Colombian sugarcane soils

Knowing chemical soil properties could determine crop management and total production. Traditional analytical tests are time-consuming, preventing farmers from promptly taking action to achieve optimal plant and soil nutrition. The estimation of chemical properties from its spectral signal, Vis–NIRS (range from 400 to 2500 nm), is a low-cost, non-invasive, non-destructive, and fast alternative. We tested and compared four regression and six classification techniques to estimate soil chemical properties using soil samples vis–NIR spectra from the highest production region of Non-Centrifugal Sugarcane (NCS) in Colombia. The regressors models were Linear Regression (LR), Support Vector Regression (SVR), Least Absolute Shrinkage and Selection Operator (LASSO), and Cubist. The classifiers were binary trees, linear and quadratic discriminant, Naive Bayes, Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Ensemble-based architectures. Estimated properties were pH, soil organic matter and exchangeable Ca, Na, K, and Mg. We compared performance in soil property prediction, both numerical values and category labels obtained from regressors and classifiers, respectively. We obtained acceptable predictions using a regression for pH (R²=0.81, ρ=0.9), OM (R²=0.37, ρ=0.63), Ca (R²=0.55, ρ=0.75), Mg (R²=0.44, ρ=0.67), confirming the predictive performance reported in the literature. Additionally, we labeled pH (Acc = 0.87), OM (Acc = 0.73), Ca (Acc = 0.72), and Na (Acc = 0.99) with acceptable accuracy. The variability of the soil matrix of the sampling area is an important and valuable limitation for the construction of models. Our results suggest that labeled soil samples and machine learning classifiers might help as a potential tool for supporting decision-making processes in soil and plant nutrition for agriculture.