ABSTRACT
Salinization in arable land using unmanned aerial vehicles, multispectral images, and machine learning

Patricia Paulina Hernández-Victoria1, Héctor Flores-Magdaleno1*, Abel Quevedo-Nolasco1, Jorge Flores-Velázquez1, and Gustavo Cruz-Cárdenas2
 
Soil salinity reduces crop productivity and promotes desertification. Traditional methods to assess soil salinity are costly and require a lot of time and effort. Salinity diagnosis could be more efficient by processing aerial images. This work aimed to predict the spatial distribution of soil salinity and sodicity as a function of electrical conductivity (EC) and the exchangeable Na percentage (ESP) in agricultural lands. Three hundred soil samples were collected randomly in the study area to determine in a laboratory the EC and the ESP. The reflectance of multispectral images obtained with a camera transported in an unmanned aerial vehicle was used to calculate several reflectance indices which were used to predict EC and ESP categories. Five automated learning algorithms: Neural Networks, Random Forest, Support Vector Machine, Nearest Neighbors, and Decision Trees were trained and validated. The results of the Random Forest model indicated an accuracy of 0.69, recall of 0.69, F1 of 0.69, and kappa of 0.39 to predict EC categories. The Support vector machine was the most satisfactory model for predicting the ESP, obtaining an accuracy of 0.73, recall of 0.65, F1 of 0.69, and kappa of 0.41. We recommend monitoring the study area at different year times, by identifying and mapping areas with salinity problems, farmers could implement more effective management practices, which could increase crop productivity.
Keywords: Artificial neural network, decision tree, random forest, soil salinity, support vector machine, unmanned aerial vehicles.
1Colegio de Postgraduados, Campus Montecillo, Departamento de Hidrociencias, Texcoco 56264, México.
2Instituto Politécnico Nacional, CIIDIR Unidad Michoacán, COFAAA, Jiquilpan, Michoacán 59510, México.
*Corresponding author (mhector@colpos.mx).