Modeling the risk of advancing sand areas using expert algorithms and artificial intelligence

Desertification is one of the major environmental, socio-economic problems in many countries of the world (Breckle, et.al., 2001). Desertification is actually called land degradation in dry, semi-arid and semi-humid areas, the effects of human activities being one of  the most important factors (David and Nicholas, 1994). Sand areas are one of the desert  landforms, whose progress and development can threaten infrastructure facilities. The timely and correct identification of the changes in the earth's surface creates a basis for a better understanding of the connections and interactions between humans and natural phenomena for better management of resources. To identify land cover changes, it is possible to use multi-temporal data and quantitative analysis of these data at different times (Lu, et.al., 2004), therefore, one of the accurate management tools that causes the application of management based on current knowledge, these studies Monitoring is done using the mentioned data. The use of satellite data and ground information in such studies has caused many temporal and spatial changes of phenomena to be well depicted, which can be beneficial in better understanding  and  interaction with the environment and ultimately its sustainable management  and development. To obtain and extract basic information, the best tool is to use telemetry technologies, which by using satellite data, in addition to reducing costs, increases accuracy and speed, and its importance is increasing day by day in the direction of sustainable development (Alavi Panah, 1385). Since field studies in the field of spatial changes of sandy areas of this city are difficult and expensive to repeat, facilities such as simulating these areas with expert algorithms and artificial intelligence can be used to investigate and monitor critical areas at regular intervals. Accurate and economically appropriate. Therefore, in this research, with the aim of investigating the effectiveness of these models in the periodic changes of the sandy plains of Ferkhes plain, two algorithms, perceptron neural network and random forest, were chosen, and the reason for choosing these models is the ability to model according to the existing uncertainties, interference Fewer users and insensitivity of the model to how the data is distributed.

The progress and development of the sandy areas of the Fern Plain depends on three factors, climatic, environmental and human. Therefore, the input variables to the expert and artificial intelligence models were chosen to cover these three factors. Therefore, factors such as drought, the number of dusty days, as well as vegetation index were entered into the model as dynamic variables, and environmental factors such as soil, elevation and altitude, geology, slope and direction were entered into the model as static variables. The statistical period investigated for the changes of wind erosion zones was considered to be 15 years from 2000 to 2015, based on this time base, qualitatively homogeneous and reconstructed meteorological data and images A satellite was selected and processed in 5-year periods (2000, 2005, 2010 and 2015). Modeling of the changes of sandy areas was done using two algorithms of perceptron neural network and random forest in MATLAB software environment. To choose the best neural network structure, a large number of neural networks with different structures were designed and evaluated. These neural networks were built and implemented by changing adjustable parameters (including transfer function, learning rule, number of middle layer, number of neurons of middle layer, number of patterns). One of the most common types of neural networks is multilayer perceptron (MLP). This network consists of an input layer, one or more hidden layers and an output. MLP can be trained by a back propagation algorithm. Typically, MLP is organized as a set of interconnected layers of input, hidden, and output artificial. The accuracy of these networks was checked by the statistical criteria calculated in the test stage, and finally the network that had the closest result to the reality was selected as the main network. The main active function used in this research is sigmoid, which is a logistic function. Then by comparing the network output and the actual output, the error value is calculated, this error is returned in the form of back propagation (BP) in the network to reset the connecting weights of the nodes (Chang and Liao, 2012). Other evaluation indices MSE, RMSE and R were used as network performance criteria in training and validation. The selection of Fern plain as a study area is due to the high potential of this area in the advancement of sand areas, for this purpose, 8 effective factors in the development of these areas were investigated. These factors were entered into the model in the form of three dynamic indices and five static indices.

In evaluating the results of modeling algorithms, dynamic variables in all periods were introduced as the most important factors in the occurrence of wind erosion and the advancement of sand areas. The diagram of the importance of predictor variables is presented in Figure 7. The results show that the vegetation cover index ranks first in all periods, the drought index ranks second in 2000 and 2015, and the dust days index ranks third in these two years. Meanwhile, in 2005 and 2010, the dust index and drought index ranked second and third respectively. Among the static variables used in this research, the height digital model variable was ranked fourth in 2000 and 2010, and in 2005 and 2015, geological and soil variables were important. In almost all studied periods, the direction factor is less important than other factors, which can be removed from the set of variables required for modeling to predict sand areas.