In this study, non-linear models were obtained through machine learning as alternative techniques for determining soil agrochemical parameters—phosphorus, potassium, nitrogen, pH and organic matter (humus) content. The models were created using statistical methods, analyzing color characteristics of digital images from different optical devices. Regression models were modeled using a Broyden-Fletcher-Goldfarb-Shanno (BFGS) neural network. The models were evaluated using evaluation criteria—coefficient of determination (R2), mean square error (MSE), root mean square error (RMSE), residual prediction deviation (RPD), mean absolute error (MAE) and mean absolute percentage error (MAPE). In the models compiled by means of neural networks for determining parameters pH, potassium, content of organic carbon-humus H and phosphorus P, average absolute percentage error values of 9.64%, 6.92%, 7.48% and 21.55% were obtained, respectively. The models have RPD values in the reference ranges of 3 to 8 and show high robustness to new inputs. The proposed tools could be used in modern smart agriculture as a complementary method for express monitoring of agrochemical soil indicators. Patterns can be implemented when programming mobile web-based applications.