Spectral analysis - artificial intelligence model with spatial variables for some soil chemical properties estimation

Introduction

The lack of scale-dependent information related to the soil properties has largely led to limitations in agriculture, hydrology, climate, ecology, and environmental studies. In this regard, the aim of study is to analyze the scale-dependent changes in estimating the spatial distribution of soil pH and salinity in Tabriz city. The proposed method applies the decomposition approach to increase the accuracy of soil properties spatial distribution estimation which includes creating a relationship between soil salinity and pH with the sub-series of input variables from the decomposition method, construction a relationship between soil salinity and pH sub-series with sub-series of input variables and finally, using a data reduction approach for the effective sub-series determination from decomposition.

Materials and Methods

The sampling points of study are related to Tabriz city in East Azerbaijan province. 60 and 38 samples were collected from 0 to 20 cm depth for soil pH and salinity monitoring, respectively. Changes in soil properties are evident from one region to another. The cyclical behavior of soil properties is called periodicity. Previous studies in the description of spatial changes have mainly focused on the spatial similarities obtained between properties in a region, which the change of spatial arrangement-spatial dependence or periodicity of soil properties has not been regarded. Spectral analysis can measure the periodicity in spatial changes by approximation a series of spatial data with the sum of sine and cosine functions. The purpose of discrete wavelet transformation is to decompose the signal into the sub-series in order to obtain a comprehensive analysis of the input signal. Discrete wavelet transform is used to calculate approximation coefficients in a signal. Maximum overlap discrete wavelet transform (MODWT) is similar to discrete wavelet transform in which low and high pass filters are applied to the input signal at each level. However, the elimination of coefficients is not done by MODWT. Principle component analysis was used as the reduction method to find the effective sub-series of decomposition.

Results and Discussion

In this study, 48 and 12 points were used in soil pH modeling, and 30 and 8 points used in soil salinity modeling in the calibration and validation periods, respectively. The average pH of the area was equal to 7.7. After determination the salinity and pH of the soil samples, modeling of spatial variables and soil properties was done. Longitude, latitude, height above sea level, slope and slope aspect were considered as the spatial variables, and the longitude, latitude and height of sampling points were recorded by GPS , but the slope and slope aspect of sampling points were taken from the production maps of slope and direction. They were extracted from the DEM map of Tabriz city. The changes in accuracy measurement indices show the varation in the estimated soil salinity and pH against different inputs of SVR model. As the level number increases in MODWT, the correlation coefficient increases. In the case of salinity, analyzing only the aspect of slope has been able to increase the accuracy of the measurement indices, the reduction rate of RMSE, RRMSE, Var from the previous optimal state without decomposition to decomposition state is 3%-4%-20%, respectively, and the rate of RPD increase is equal to 4%. Based on the values of the accuracy measurement indices, principal component analysis can increase the accuracy of the estimated values, for example, the amount of reduction in RMSE using principal component analysis compared to the MODWT decomposition in pH and salinity was 4% and 2%, respectively. The box plot of data related to the use of principal component analysis has become closer to the box plot of the observation data compared to the case where only the decomposition method was used. This issue is quite evident in the case of Whikeres lines (areas related to the first and third quadrants). In the drawn dendrogram, the distance of data using principal component analysis to the observational data decreased, and this problem shows the increase in accuracy with combination of MODWT spectral approach and the PCA data reduction method.



Conclusion

In recent years, there is an increasing demand for soil spatial distribution information in environmental decision-making and land use management. One of the issues that can affect the accuracy of spatial modeling of soil properties is the spatial information increase of input variables to the model. Modeling sub-series of input variables with sub-series of soil properties and finally the sum of estimated sub-series of soil properties could not increase the accuracy of estimated values. Decomposition of some series could increase the accuracy of estimates. The factors that can affect the accuracy of the proposed method for determination the spatial changes of soil properties include the type of input variables, the type of used model in the modeling process, the use of the appropriate method in spectral analysis and determining the effective factors in the decomposition method. Therefore, the combination of spectral analysis and artificial intelligence as an effective option can increase the accuracy of the spatial distribution of soil properties.