This paper describes the finding by machine learning, based on artificial neural networks, non-linear models for determining the values of the most frequently observed agrochemical parameters of the soil—content of organic matter (humus), pH, phosphorus, potassium and nitrogen. The models were obtained using statistical methods for analyzing color characteristics of digital images from various optical devices. The results obtained from the authors’ previous research on finding regression models with a single regression factor are not accurate enough to be used in practice. In the present study, four regression factors are used, which are different for each optical device. Machine learning with Broden-Fletcher-Goldfarb-Shanno (BFGS) neural networks was applied. Mean absolute error (MAE), mean absolute percentage error (MAPE), mean square error (MSE), root mean square error (RMSE), residual prediction deviation (RPD) and coefficient of determination (R2) were used to evaluate the models. In the models compiled by neural networks for determining indicators pH, potassium, content of organic carbon-humus H and phosphorus P, MAPE values were obtained, respectively 6.91%, 6.92%, 7.39% and 20.32%. The models show RPD values in the reference interval of 3 to 8, which proves their high robustness to new input data. Modern smart agriculture can use the proposed approach for express monitoring of agrochemical indicators of the soil, including in the field. Models can be embedded in mobile web-based applications.