Comparison of artificial neural network and multivariate regression models for prediction of Azotobacteria population in soil under different land uses

- The study describes relationships between soil physicochemical and biological parameters and Azotobacter population in soil (enumerated in LG and WG media) of agricultural and non-agricultural lands from the East Azerbaijan, Ardabil and Gilan Provinces, Iran.
- We observed negative relationships between Azotobacter number in both media with EC, silt and CCE of soil. In addition, we found positive relationships among Azotobacter population with SIR, bacteria population and sand content of soil.
- Estimation of Azotobacter number by ANN and regression models with other measured soil parameters was main objective of this research. According to the result we presented an ANN for estimating Azotobacter in WG an LG by using 5 variables of soil as inputs of the model. These inputs were soil texture (including sand and silt), soil salinity, SIR, and content of CCE in soil.
- Comparison of the studied models showed that Azotobacter population on WG medium better than LG medium can be estimated by ANN.
- An assessment of the importance of individual input parameters has indicated that factors influencing on soil Azotobacter number including EC, pH, OC, soil texture and carbonate calcium.

Azotobacteria are one of the most important and beneficial soil bacteria which their number and distribution are affected by physicochemical and biological properties of soil and land usage. The aim of this study was to evaluate the population of Azotobacter in soils with different land uses and relationship between population size and some physicochemical and biological properties of soil by using artificial neural network (ANN) and multivariate linear regression (MLR) methods. In total, 50 soil samples were collected from depth (0-25 cm) under different land uses located in East Azerbaijan, Ardabil and Gilan provinces, Iran. Population of Azotobacter was separately counted in Winogradsky and LG media by preparation of serial dilution and plate counts. In addition, soil texture, pH, electrical conductivity (EC), carbonate calcium equivalent (CCE), organic carbon (TOC), cold water extractable OC (CWEOC), hot water extractable OC (HWEOC), light fraction OC (LFOC), heavy fraction OC (HFOC), basal respiration (BR) and substrate induced respiration (SIR), the number of bacteria, fungi and actinomycete were measured in three replicates in each soil sample. To predict Azotobacteria population based on easily measurable characteristics of soil properties, MLR analysis and ANN model (feed-forward back propagation network) were used. In order to assess the models, root mean square error (RMSE) and R2 were used. The R2 and RMSE values for population of Azotobacter in Winogradsky medium obtained by ANN model with SIR, EC, CCE, sand and silt as entered variables were 0.76 and 0.36, respectively, and for population of Azotobacter in LG medium, were 0.45 and 0.50, respectively. Using MLR the R2 value for population of Azotobacter in WG and LG media was 0.63 and 0.39, respectively. Results showed that ANN with eight neurons in hidden layer had better performance in predicting population of Azotobacter in WG than MLR.