This paper describes a plant complex and development of the karst spring Rakonek as an important resource for water delivery of south Istria which is used by Vodovod Pula. Considering the connection of the spring with the basin of undercurrents of river Pazinčica in central Istria, general features of the entire area have been laid out and their interconnection. Along with systematization and consolidation of available registered data, a hydrological analysis of the water regime has been done and after that an analysis of the turbidity of the spring waters with the useful insight in mind and getting the necessary data for evaluation and prognosis of critical states. In the analysis of the data a twenty-fiveyear span has been taken into account from which basic statistical parameters, characteristically values and situations have been drawn along with analysis of frequency, durability and probability. In such a manner a summary of rainfall regime and water level along with amount of drawn, overflow and total water flow on the spring. Methodological representation and results of the analysis of the water regime and turbidity, besides being founded on basic statistical analysis, include methods of deep data analysis with the application of artificial intelligence. Basic steps needed for understanding the Weka system for machine learning have also been laid out with the immediate application in order to identify the appearance of turbidity on the spring. The attributes have been ranked on importance order and their influence on the turbidity, and considering the ordering of the attributes we could immediately remove those attributes that have no or little significance which contributed to the model accuracy and overall quality. Two classificators have been used: regression decision tree and multilayered neuron nets. We have observed their adaptability to the registered values. Modeling has been carried out in two phases: training on the learning data and verification of the model on an independent data sequence. Verification was carried out in such a way that model trained on a larger part of independent data and then tested the results on the rest. Analytical indicators and graphical models needed for interpretation were obtained in such a way. Considering that the chosen model has been obtained with the regression decision tree classificator the system allows for its graphical visualization. The results of the prognosis of the increased turbidity are displayed in percentage of the test data sequences, and the adaptability of the aforementioned is displayed with line graphs.