<h2>DESCRIPTION</h2>
<p><em>r.landslide</em> allows the identification of landslide susceptible areas on the basis of Artificial Neural Networks (ANNs), environmental parameters and landslide databases.</p>
<h3>ANN parameters tab</h3>
<p>In this guide, the user must enter the parameters related to the ANN architecture. There is the option to produce the susceptibility map only with an ANN, or from the ANN that had the best performance within a set of networks, from batch mode option (To do this, the user must check the box 'Train a set of ANNs and select the best one [batch mode]').</p>
<p><strong>Single ANN mode</strong></p>
<p>To create the map of susceptibility from a single ANN architeture, the user <strong>should not</strong> check the box 'Train a set of ANNs and select the best one [batch mode]'. The parameters required to use this mode are:</p>
<p><strong>Number of hidden neurons: </strong>The number of neurons in the hidden layer must be an integer. It is recommended that the maximum number of neurons in the intermediate layer be defined based on the Netch-Nielsen equation (Netch-Nielsen, 1987), which ensures that the neural network is able to approximate any continuous function and that depends on the number of thematic input parameters used by the user: N<sub>H </sub>&le; 2N<sub>I</sub> + 1,&nbsp;where&nbsp;N<sub>H</sub> is the number of neurons in the intermediate layer and&nbsp;N<sub>I</sub> is the number of input parameters. A default value of 12 is set.</p>
<p><strong>Learning rate: </strong>Learning rate is a hyper-parameter that controls how much we are adjusting the weights of our network with respect the loss gradient.&nbsp;This paramter is generally chosen in the interval [0,1]. Lower learning rates may require a great number of training epochs. A default value of 0.6 is set.</p>
<p><strong>Number of epochs: </strong>An epoch is a measure of the number of times all of the training vectors are used to update the weights. A default value of 200 is set. In order to avoid overfitting due to a probable high number of epochs, the early stopping method was implemented.</p>
<p><strong>Batch mode</strong></p>
<p>Batch mode is activated when the 'Train a set of ANNs and select the best one [batch mode]' box <strong>is selected</strong>. In this mode, a set of ANNs is generated from the variation of the number of neurons in the hidden layer and the set of initial conditions of the synaptic weights. The network that presented the best performance is then used to produce the susceptibility map.</p>
<p>To do this, the user must specify the parameters marked with [batch mode] in the tab, in addition to the learning rate and number of epochs parameters previously described.</p>
<p><strong>Minimum number of hidden neurons: </strong>Minimum number of hidden neurons to be tested. It is recommended that this parameter is not less than 2.</p>
<p><strong>Maximum number of hidden neurons: </strong>Maximum number of hidden neurons to be tested. As previously mentioned, it is recommended that this value does not exceed that proposed by the equation of Netch-Nielsen&nbsp;(Netch-Nielsen, 1987) in order to ensure that the neural network is able to approximate any continuous function. If this parameter is set to 0, the maximum number of neurons will be calculated automatically based on equation.</p>
<p><strong>Number of initial conditions: </strong>Number of initial sets of synaptic weights that will be tested for each set of neurons in the hidden layer. Attention: too high values can result in a long processing time.</p>
<p><strong>Optional parameters of the ANN parameters tab</strong></p>
<p>Considering that the construction phase of an ANN comprises three main steps: training, validation and testing, the input database must be restructured into three sets to serve these processes. The module allows the user to define a percentage of data that will be selected for the validation step and for the test step (0-1); however, it is recommended that the default values of 0.15 be maintained to avoid losses in the learning process from ANN.</p>
<h3>Training tab</h3>
<p>In this tab, the raster format thematic input parameters and the landslides and non-landslides events location vector files must be inserted. In this tab, there is also the option to perform only the training, validation and test steps. In this case, the module does not perform the calculation of susceptibility maps. This allows different network structures to be tested before the final susceptibility of the study area is evaluated. To do this, the users have to check the box 'Perform ONLY the training, validation and test steps'.</p>
<p>A description of the parameters is given below:</p>
<p><strong>Thematic layers: </strong>Thematic layers refer to environmental parameters of the region study that may influence the occurrence of a landslide event. Some examples include: elevation, slope, aspect, topographic wetness index, soil type, soil use, among others. The user can add different layers, and can verify the influence of each one of them by the sensitivity analysis performed by the module.</p>
<p>This information can be obtained from online databases for the region of interest and must be entered into the module in raster format.</p>
<p><strong>Point vector with landslides locations: </strong>Here, the user must indicate the vector file containing the location, in point format, of places where there were landslides of the region of interest. It is from this georeferenced information and thematic layers that the module will create the database for ANN training and application. It is recommended that for good ANN learning, a considerable number of slip points should be entered (averaging &gt; 50, depending on the size of the area).</p>
<p>This type of information can be obtained from online databases, scientific works, public organs, etc.</p>
<p><strong>Point vector with non-landslides locations: </strong>The user must specify the location of examples of places where there was no occurrence of slides in point vector format. The generation of these localities is usually given in a random manner, and should not overlap the landslide regions. It is recommended to generate the same number of points present in the vector file of landslide points.</p>
<p><strong>Text-file with X,Y coordinates of landslides location:</strong> This is an <strong>optional</strong> feature that allows the user to enter a text file containing the coordinates in the form X, Y that will be automatically converted to vector format, instead of entering the vector file of the location of landslides from the previous option. The user can also enter the values directly in the box that appears.</p>
<p><strong>Name of output susceptibility map: </strong>If the training-only checkbox is not checked, you must enter the name of the output raster map for susceptibility.</p>
<p><strong>Directory name to save files (it will be created): </strong>You must enter the path to create the directory that will save performance results, ANN architecture, and sensitivity assessment.<strong><br /></strong></p>
<h3>Reckoning tab</h3>
<p>The Reckoning tab should be used in the case where only the network training, validation and test steps were performed by checking the 'check box' in the Training tab specified above. To perform this process, it is necessary to check the 'Perform ONLY the reckoning' check box and insert the directory containing the ANN architecture parameters which were stored after the training. Also, the user must identify the name of the map of susceptibility that will be generated.</p>
<h3>Optional tab</h3>
<p>Here, the user can insert a raster layer to delimit a smaller region of interest for the generation of the susceptibility map. When this option is used, a temporary region is created from the active layer. This function is interesting in cases where the user wants to perform the evaluation of a smaller area, thus reducing time and computational demand.</p>
<h3>Outputs</h3>
<p>The module will generate a raster map representing the susceptibility to landslides of the area of interest. The raster will contain values ranging from 0 to 1, values close to 0 will represent areas of lesser susceptibility, while values closer to 1 will represent areas of greater susceptibility to these events.</p>
<p>In the user-specified directory on the Training tab, different information about the selected ANN performance will be saved, which include: training error graph, validation and test, text files containing input information, ANN architecture, and performance, and graphs of the sensitivity evaluation of each of the parameters used.</p>
<p>Note that the directory still includes text files called weights, peights, biasH and biasO, which will be used by the Reckoning tab. Therefore, it is recommended that the names of these files not be changed.</p>
<h2>NOTE</h2>
<p>After generating the module's susceptibility map, the user must load the raster map to the workspace via the <em>d.rast</em> module.</p>
<h2>REFERENCE</h2>
<p>R. Netch-Nielsen, Kolmogorov&rsquo;s mapping neural network existence theorem, First IEEE International Joint Conference on Neural Networks 26 (1987) 11&ndash;14.</p>
<h2>SEE ALSO</h2>
<p><em> <a href="http://grass.osgeo.org/programming7/">GRASS Programmer's Manual</a> </em></p>
<h2>AUTHOR</h2>
<p>Lucimara Bragagnolo</p>
<p>Roberto Valmir da Silva</p>
<p>Jos&eacute; Mario Vicensi Grzybowski</p>
<p><em>Last changed: $Date: 2019-01-03 12:59:19 +0200 (Thu, 03 Jan 2019) $</em></p>