EFFICIENCY OF MACHINE LEARNING ALGORITHMS IN SOIL SALINITY DETECTION USING LANDSAT-8 OLI IMAGERY

Abstract

Climate change is one of the biggest problems facing today's world. Rising temperatures and declining rainfall have had a profound effect on the planet, one of which is the destructive effects of soil salinity. Soil salinity phenomena commonly occur in arid and semi-arid regions. Maharloo Salt Lake, southeast of Shiraz, Iran, with an arid and semi-arid climate, has faced severe droughts in the past and is dealing with the soil salinity problem. One useful way to manage land and soil in such areas is regular monitoring of the soils and lands and keeping abreast of changes to prevent land degradation and erosion. With the advancement of technology, remote sensing techniques to monitor natural factors have become very popular. Landsat sensor images were used in this research, and several environmental indicators were extracted by combining satellite bands. Three machine learning algorithms, RF, GBM, and MLP, were used to evaluate methods for monitoring and mapping saline soils. The models were trained and then tested to compare the accuracy and performance of each model in predicting soil salinity. GBM algorithm showed the best performance with R2 = 0.89 and RMSE = 0.63 for testing the dataset after that RF model with R2 = 0.85 and RMSE = 0.71 and the worst performance was for MLP model with R2 = 0.75 and RMSE = 0.88. The figures mapped from the results of these algorithms for salinity distribution in this region showed that by choosing the appropriate algorithm and suitable in-situ data, it could be possible to estimate soil salinity to an excellent extent by satellite data.