رضا عموزاده

Due to the importance of temperature and salinity parameters of rivers, seas and dams due to their impact on the ecosystem and the quality of water resources, as well as the progress and preva‎lence of using soft computing methods in various sciences, including oceanography, and the advantage of using it in modeling due to time consumption and lower cost while maintaining accuracy in the results, using machine learning algorithms for predicting salinity and water temperature in the Persian Gulf using the data of the salinity-temperature-depth device have been chosen as the goal of the research. In this research, the field data of the salinity-temperature-depth device obtained from different cruises as model outputs and according to the geographic location and time of data collection, various combinations of information on water depth, wind speed, precipitation, evaporation, net heat flux, temperature and water surface salinity, tidal velocity and amplitude are selected as model inputs. Two methods of multilayer perceptron (a type of machine learning algorithm) and H2O autoML framework (containing a set of machine learning algorithms) were applied to the inputs and outputs of the dataset in order to estimate the mentioned properties. In the research that was conducted, among the different conditions assumed for the effective cases in the temperature profile and salinity in the Persian Gulf, the analysis of the results showed that the variables of latitude and longitude, day of the year, water surface temperature and wind speed in the estimation of the temperature profile and also the above variables plus tidal velocity have had a greater impact on the estimation of the salinity profile in the Persian Gulf than other competitors (correlation coefficients of 0.93 and 0.94 and the root mean square error of 0.35℃ and 0.11 psu between the actual and predicted values in the test data, respectively, for temperature and Salinity at 15 levels from the surface to a depth of 50 meters). Among the scenarios that were considered for the estimation of the salinity profile, the modeling results of 925 data sets without considering surface salinity as one of the inputs are better than the modeling of 331 data sets with the same inputs plus considering surface salinity as one of the inputs, which shows the effect of the dataset volume in the generalization of the model. Examining the correlation coefficients of the top models in different time situations shows that the temperature and salinity profiles are estimated more accurately in the winter season. Also, by comparing from the point of view of location, it has been concluded that both temperature and salinity profiles in the longitude range of 52-57°E (east of the Persian Gulf and Strait of Hormuz) have been more accurate compared to the Gulf of Oman and the west of the Persian Gulf. Examining the accuracy and errors showed that the H2O autoML framework is a powerful tool in solving problems with a small amount of data compared to deep learning methods that often require large datasets.

نيم‌رخ شوري-دما , خليج فارس , دستگاه شوري-دما-عمق , محاسبات نرم , هوش مصنوعي

Temperature , persian gulf , conductivity-temperature-depth (CTD) , soft-computing , Artificial Intelligence (AI)

Reza Amoozade

Dr. Mostafa Siyadat