Determination of the Most Important Factors on Rainfed Wheat Yield by Using Sensitivity Analysis in Central Zagros

Wheat (Triticum aestivum L.) as the most strategic crop for human nutrition is cultivated in many countries under rainfed conditions in semiarid regions. To be vital importance to predict rainfed wheat yield and determine the important factors which affect this crop. Modeling is one of the approaches to predict the response of land to land use. Artificial neural networks (ANN) are considered as one of the modeling approaches to yield prediction and determination of the most important parameters in crop productions. In rainfed wheat hilly land of central Zagros of Iran, there are various parameters that influence this crop production. Therefore, the objective of this study was to identify these important factors.
This study was conducted for two years and at two sites under rainfed conditions in Koohrang and Ardal districts in Chaharmahal and Bakhtiari provinces, central Zagros of Iran. At both sites, the study was made on farmer–operated winter wheat fields. At the Koohrang and Ardal sites, 102 and 100 sampling points were selected, respectively. 202 sampling points were chosen on the landscape covering summit, shoulder, backslope, footslope, and toeslope at two sites with varying climatic conditions. Four parameter groups including terrain attributes, soil physical and chemical properties, precipitation, and weed biomass, including 54 factors were used as the inputs, and wheat grain and biomass yield as the targets for ANN models. A feed-forward back-propagating ANN structure was used to develop yield prediction models. The data set was randomly shuffled; 60%, 20% and 20% of them were used for the learning network, testing and verification, respectively. After determination of the best structure of ANN model, crop yields were predicted by the ANN models. By the Hill sensitivity analysis method (Hill, 1998), response of each factor was studied and determined the most effective parameters on grain and biomass yield. This method calculates relative sensitivity coefficient by dividing the sensitivity coefficient of every variable when the variable is reduced 10% by the maximum sensitivity coefficient, therefore the maximum relative sensitivity coefficient is 1.
The descriptive statistics for various soil characteristics showed that, soil chemical and physical parameters can be classified into three orders. Sand, TN, Kava., Pava., SOM, CCE, and gravel showed high variability (CV>35); clay, silt, and CEC had moderate variability (3515); and SP and pH indicated low variability (CV