Soil cone index prediction using artificial neural networks model and its comparison with regression models

Soil cone index as one of the criteria that states mechanical strength of the soil is affected by many factors including soil moisture content and soil compaction. Despite widespread progress in the development of precision agriculture, crop production management in relation to soil physical properties goes back to the last few years. One of the most important soil characteristics that affects crop yield is soil mechanical resistance. This characteristic is expressed generally by soil cone index maps. In this study, for measuring and determining the factors affecting soil cone index, field experiments were carried out on three soil types. Within each soil type, the factorial experiment based on randomized complete block design (RCDB) with five replications was used. The effects of soil moisture content at three levels (dry, semi-humid and humid), sampling depth at three levels (0-10, 10-20 and 20-30 cm) and number of tractor traffic at three levels (0, 10 and 20 Time passes) was investigated on soil cone index. After data analysis, it was revealed that the effects of soil type, sampling depth, different levels of soil moisture and tractor traffic were significant on soil cone index values (P<0.01). In order to develop a mathematical model for soil cone index, multivariate linear regression was used. Independent factors were soil moisture content, soil bulk density, electrical conductivity and sampling depth whereas soil cone index was the dependent factor. The results showed that the effect of all independent variables on soil cone index as the dependent variable were significant at probability level of 1%. Back propagation neural networks with three different training algorithms (gradient descending algorithm with momentum, descending scaled gradient and Levenberg-Marquardt) were adopted for predicting soil cone index parameter. Back propagation neural networks with Levenberg- Marquardt training algorithm presented better accuracy in simulation and prediction as compared to others. Using Levenberg-Marquardt training algorithm with two hidden layers with 34 neurons in each layer presented the best performance than other algorithms and even Levenberg Marquardt training algorithm with one layer. Comparison of results of artificial neural network models and regression models to predict the soil cone index indicated that the neural network model could model soil cone index values with higher accuracy than the regression models. The results of this study could be utilized in soil compaction management of Ardabil plain soils and also in determination of the optimum tillage depth in these areas.