Sunum Açık Erişim

Spatial predictability of salinity hazard with machine learning algorithms and digital data in the Irrigation Plain

2021    Kaya, Fuat; Başayiğit, Levent

Introduction, scope and main objectives

In the irrigation area in semi-arid regions, spatially detection studies are important marker for accurate monitoring of soil salinization. As a result of irrigation, capillarity, which is the soil dynamic system, tends to increase soil salinity on the surface (Scanlon et al, 2016; Hopmans et al., 2021). The association of socioecological differences with irrigation and the use of land cover classes as a variable that can reflect its differences (Mponela et al., 2020) are investigated. Machine learningbased modeling approaches in agricultural systems are used to suggest decision/support systems (Yamaç, 2021).

Methodology

The study was carried out in an agricultural plain where irrigation activities have been carried out for a quarter of a century. EC (µS/cm) was determined in the saturation paste (Rhoades, Chanduvi and Lesch, 1999) for 91 samples taken from the field. Environmental variables were generated from Sentinel 2A-MSI satellite, Digital Elevation Model, and CORINE Land Cover Classes. The data set was divided into 70 percent training and 30 percent test set. Relevant packages were used in R Core Environment in data set preparation processes and modeling (R Core Team, 2021; Omuto et al., 2020). Ordinary kriging was applied by controlling the normal distribution of the dependent variable. Also, random forest algorithm spatial modeling was used. In the hybrid (RF-Regression Kriging) approach, explanatory variation is estimated by RF algorithms and the process is carried out by summing the regression value of EC and the kriging values of model residuals in non-sampled locations. Root mean square error (RMSE) values were used as model accuracy criteria.

Results

The mean of EC values, standard deviation, coefficient of variation, minimum and maximum were 612.1, 288.5, 47.1 (%), 110.0, and 2068.0, respectively. Permanently Irrigated Land Class has the highest average EC value (828.9 µS cm-1). Modeling and cross-validation resulted from ordinary kriging, the RMSE value was determined 270.8 µS cm-1. For the RF model, the RMSE value was determined at 102.4 µS cm-1 in the training set and 314.0 µS/cm in the test set. The most important environmental variables in the random forest model were CLCC 212 Permanently Irrigated Land, Aspect, and Normalized difference vegetation index. For the RF-RK approach, the RMSE value approached zero in the training (RMSE: 5.9e-14 µS/cm) and test sets (RMSE: 9.6e-14 µS/cm).

Discussion

The approach of machine learning-based modeling was given relatively accurate results compared to the geostatistical-based modeling, furthermore, the hybrid modeling technique obtained more accurate modeling results than both approaches. In machine learning-based modeling approaches, the location of sample points can also be neglected (Hengl et al., 2018). Land cover class and NDVI value were found to be important factors in the random forest model, indicating that agricultural activities carried out on the land are also important for salinity risk (Maleki et al. 2020).

Conclusions

Machine learning-based modeling approaches using land cover classes as environmental variables may be preferred to mapping soil salinity using a purely geostatistical method. In hybrid modeling approaches, the spatial relationship present in model residuals significantly improves model accuracy. It can provide more accurate insights into salinity management and monitoring from digital maps produced. The views expressed in this information product are those of the author(s) and do not necessarily reflect the views or policies of FAO.

References

Scanlon, B.R., Reedy, R.C., Xu, P., Engle, M., Nicot, J.P., Yoxtheimer, D., Yang, Q. et al. 2020. Can we beneficially reuse produced water from oil and gas extraction in the U.S.? Science of The Total Environment, 717: 137085. http://www.fao.org/3/cb4809en/cb4809en.pdf Hengl, T., Nussbaum, M., Wright, M.N., Heuvelink, G.B.M. & Gräler, B. 2018. Random forest as a generic framework for predictive modeling of spatial and spatio-temporal variables. PeerJ, 6: e5518. https://doi.org/10.7717/peerj.5518 Hopmans, J.W., Qureshi, A.S., Kisekka, I., Munns, R., Grattan, S.R., Rengasamy, P., Ben-Gal, A. et al. 2021. Chapter One - Critical knowledge gaps and research priorities in global soil salinity. In D.L. Sparks, ed. Advances in Agronomy, pp. 1–191. Academic Press. https://doi.org/10.1016/bs.agron.2021.03.001 Maleki, S., Karimi, A., Zeraatpisheh, M., Poozeshi, R. & Feizi, H. 2021. Long-term cultivation effects on soil properties variations in different landforms in an arid region of eastern Iran. CATENA, 206: 105465. https://doi.org/10.1016/j.catena.2021.105465 Mponela, P., Snapp, S., Villamor, G., Tamene, L., Le, Q.B. & Borgemeister, C. 2020. Digital soil mapping of nitrogen, phosphorus, potassium, organic carbon and their crop response thresholds in smallholder managed escarpments of Malawi. Applied Geography, 124: 102299. https://doi.org/10.1016/j.apgeog.2020.102299 Omuto, C.T., Vargas, R.R., El Mobarak, A.M., Mohamed, N., Viatkin, K. & Yigini, Y. 2020. Mapping of salt-affected soils – Technical manual. Rome, Italy, FAO. 112 pp. https://doi.org/10.4060/ca9215en R Core Team. 2021. R: The R Project for Statistical Computing [online]. [Cited 27 Aug 2021]. https://www.r-project.org/ Rhoades, J.D., F, C., Lesch, S., Smith, M., Organization, (FAO) Food and Agriculture. 1999. Soil salinity assessment : methods and interpretations of electrical conductivity measurements. FAO irrigation and drainage paper. Rome, FAO. 149 p. p. (also available at http://www.fao.org/3/x2002e/x2002e.pdf). Yamaç, S.S. 2021. Analysis and modeling of agricultural systems. In B. Pakdemirli, H.Ö. Sivritepe, Z. Bayraktar, S. Takmaz, Eds. Next Generation Agriculture after the Pandemic, pp. 149–170. Ankara, TR, Sonçağ publications. (In Turkish).