Abstarct: Due to spectral diversity, duplication covering and low cost, the remote sensing in comparison to the rest of information gathering methods has lower cost and more effective. Now a days, these data are the primary means for studying the Earth's surface and its constituent factors. The possibility of digitization of data has led computer systems to use these data without intermediaries. The prominent application of remote sensing is planning land cover maps. Land cover maps, along showing the capacities of the land in the current state, provide useful information about the land cover situation for experts so far. However, these data depend upon the time of preparation. In this way, the process of destruction of environmental resources in speperiods of time can be examined. Nevertheless, due to the huge volumes of remote sensing data, the processing by the traditional architectures for information processing software is very complicated and time consuming. To cope with this problem in this thesis, an architecture for processing big remote sensing data baxsed on cluster computing is presented. On the other hand, an autonomous technique is explored for the production of a high resolution land cover map. In the proposed architecture, by preprocessing the data, a reference library by the optimum combination of the spectral features is created. So then, the learning model is taught using random forest. Finally, the proposed model is used to distinguish nine predefined classes of land cover and to provide a map of Iran's coverage of over 30 meters resolution. Landsat 8 satellite imagery is used for training process. Each Landsat 8 image band includes 62 million pixels. Five known satellite image scenes are selected for testing and results show significance improvement versus traditional architectures. Aforementioned technique with respect to time complexity is more efficient about 3.5 times.