The number of hidden layers between input and output layers directly affects the performance of the model (Basheer and Hajmeer, 2000; Karsoliya, 2012). In this study, 7 units in the geological map, 6 categories of categorical land use, DEM and derivative maps (slope, curvature, roughness, plan curvature, profile curvature, TWI, fault and river distance) were evaluated as continuous data, and a total of 21 ANN input layer was created. The dataset of the Büyük menders basin was divided into 3 as 70% analysis, 15% test and 15% validation. The hidden layer numbers varying between two-thirds and two times of the input layer number suggested by Karsoliya (2012) were tested, and result susceptibility analyzes were performed with 20 hidden layers where the best result was obtained. The training algorithm was implemented using the scaled conjugate gradient method for backpropagation and the log-sigmoid transfer function. Considering the root mean square error value in the model, the cross-entropy performance function was used. The landslide susceptibility assessment (Fig. 9a) was evaluated in five categories as very low (36.87%), low (20.79%), medium (16.34%), high (12.65%) and very high (13.35%) (Fig. 9d). According to the success- and prediction- rate curve of the obtained landslide susceptibility map (Fig. 9a, b, c,d), it is seen that 26 % of the study area is in very high group ranges and that 82.01 % of the landslides are also in these groups (Fig. 9d).
The obtained landslide susceptibility map performance and accuracy evaluations were evaluated by many different methods. The blue bars in the error histogram represent the training data, whereas the green bars show the validation data and the red bars show the test data. The histogram is an indicator of the outliers which are the data points where the mismatch is significantly worse than the majority of the data. In the artificial neural network model obtained in this study, no data mismatch between the variables and landslides was found as a result of the error histogram (Fig. 10).
Considering the root mean square error value in the model, the cross-entropy performance function was used. Cross entropy loss or log loss measures the performance of a classification model whose output layer has a probability value between 0 and 1 and is used in the ANN model. Cross entropy loss increases as the predicted probability moves away from the original value. The performance evaluation with cross entropy was calculated as a probability value of 0.038731 in the model network architecture. The analysis captured the strongest probability in the test sets in Epoch 64 (Fig. 11).
When the landslide susceptibility assessment analysis, test, validation and performance evaluations of the landslide sets in the study area are examined, it is seen that the analysis, test and validation of the landslides are estimated to be above 90% (Fig. 12). The value under the receiver operating characteristic curve of the landslide susceptibility map was calculated as 0.82, 0.84, 0.86 and 0.82 in the analysis, test, validation landslides and in the whole study area, respectively (Fig. 12). According to the accuracy tests of the landslide susceptibility map obtained as a result of the study, it is seen that the model applied in the study has a high predictive power and accuracy.