Predicting of Soil Physicochemical Properties and Productivity for Sandy Soil by Using Deep Learning Under Salt Stress Conditions

Deep learning is an exciting discipline that has already transformed the way data are analyzed in many fields. This study developed and evaluated artificial neural network (ANN), a type of deep learning algorithm, as a new way to predict the physicochemical properties of sandy soil incorporated with three rates (10, 15 and 20 ton/fed) of farmyard manure (FYM) and compost (COM)] for each treatment with three salinity levels of irrigation water 1000, 1500 and 2000 ppm. These properties were soil bulk density (Bd), available water (AW), cation exchange capacity (CEC), sodium adsorption ratio (SAR), spinach productivity (Pro). Multilayer feedforward ANN with 6 neurons in input layer and 5 neurons in output layer was trained using a back propagation learning algorithm. The ANN model was trained with data collected from previous literatures 555 observations (447 observations for training and 108 observations for testing). The model inputs were [sand, silt, clay, FYM, COM, Ec of irrigation water (ECir)]. Verification of the ANN model in prediction was done using field experimental data which carried out in Ismailia governorate (Data that an ANN model has never seen before). In order to evaluate the ANN model, root mean square error (RMSE) and correlation coefficient (R2) were calculated. After careful and extensive training, validation and testing for the ANN model were conducted. The RMSE between measured and predicted values for both Bd, AW, CEC, SAR and Pro were 0.00372 Mg.m-3, 0.166 %, 0.09903 Cmol.kg-1, 0.05975 and 12.63481 kg/fed. The R2 values were equal to 0.99835, 0.9977, 0.99765, 0.99929 and 0.99916, respectively. The high correlation coefficient for parameters outputs recall indicate for excellent prediction of ANN model for the data has never seen before.