The Matching Rate of Maps Obtained by Machine Learning Methods and Kriging Estimator in Salinity Monitoring of a Part of the Marginal Lands of Sirjan Playa, Kerman Province

Satellite images and remote sensing approaches are important tools for evaluating, mapping, and managing saline lands in different world regions. The main aim of the present study was to investigate the degree of concordance between maps obtained by machine learning methods and kriging estimator about salinity monitoring of a part of the soils of marginal lands of Sirjan Playa in two seasons, i.e., winter and summer, using two remote sensing data sources, i.e., Landsat 8 and Sentinel 2. Ninety surface soil samples (zero to 30 cm) were collected as a regular grid sampling pattern with 750 meters intervals. Some of their most important physical and chemical characteristics were determined using standard measurement methods. After performing radiometric and atmospheric corrections on mentioned satellite images, in addition to the main bands, 13 salinity indices were used to estimate soil salinity using artificial neural network, decision tree, random forest, and support vector machine models. Besides, kriging maps of soil salinity were drawn for both mentioned times. Results showed a higher performance (R2 =0.87 versus= 0.72) of Sentinel-2 than Landsat-8 in predicting soil salinity.  Moreover, results confirmed that the ANN model developed by Sentinel-2A image had the highest performance (R2 =0.77, RMSE% =27.1) to predict ECe in the winter season. Furthermore, RF presents the lowest error (R2 =0.87, RMSE% =17.4) for prediction ECe during the summer season. Among the studied salinity indices, VSSI index was also selected as the most effective index to estimate soil salinity of region. The results also showed that ECe maps obtained by two methods had a high level of concordance and an overal accuracy of over 80%; however, the change of season and type of satellite affectedthe compatibility of the maps.