Random forest model to identify changes in micro-landforms using UAV images(Case study: Afjeh region in Jajroud basin 1397-1397)

Earth is an active element in the human environment and has an important impact on other elements and general environmental characteristics. Identification and classification of land features are very important for the environment, hydrological research, geological structure analysis and other geological research (Liu et al., 2015). The land surface consists of various landforms formed by internal and external processes (Salz et al., 2016). ). Therefore, landform classification is one of the most important methods in geomorphological mapping to better understand the surface processes (Boko, 2001). Dynamic changes in different periods and existing conditions such as climate change and human activities act differently and the difference of these changes in different periods has changed the behaviors and shape of landforms, so to plan and create different models and algorithms. It is necessary to identify and measure the behavior of dynamic changes and, consequently, to identify the evolution of microlandforms. The diversity of existing UV systems is increasing rapidly and is a powerful tool in geomorphology in detecting topographic shifts to link processes and dispersions to erosion and sedimentation rates and patterns (Jeans 2012). The use of UAVs allows us to do this. Allows us to achieve the desired results by using data with a resolution of at least 10 cm, 3 bands, wide coverage and short visits (Dulrents, 2019). Recently, new algorithms and convolutional neural networks (CNNs) have become a new way of calculating several problems, such as image classification, object recognition, semantic segmentation, and so on. CNN can display a hierarchical image of the input data with acceptable resolution of the desired features. In recent years, these types of neural networks have been used to solve several tasks from different fields, including remote sensing, in which they have great advantages due to various problems such as pixel-based classification for land use, target detection such as landforms. Roads have shown human and natural features and image sharpness (Dolrenets, 2019).

In this study, the boundaries of the study area were limited using 1: 50000 topographic maps, Google Earth images, and UAV images. During a field visit to the area in the fall of 1398. UAV images were processed with ENVI 1.5 software, Arc GIS 3.10 and coded in Python program. The spatial resolution of these 10 cm images related to the two periods of 1396 and 1397 was prepared by the Ministry of Energy. Using UAV images and executing a random forest model with a map scale of 1.1000, the classification of microlandforms and their transformations were extracted. The steps and implementation method of the random forest algorithm are described below.The method used in this study to extract microlandforms is pre-trained transfer networks.In this method, the coefficients and weights trained to very deep neural networks were extracted using a UAV data set of a site with two images in the period of 1396 and 1397. In this method, very deep networks were used on this trained data in a ready-made manner with specific coefficients.

Classification based on stochastic forest modelIn order to study the changes of different microlandforms, the random forest algorithm was used and the UAV images were classified into five classes of road, bedrock rocks, soil, vegetation and water with a scale (1: 1000).g. A supervised learning algorithm was implemented on UAV images (1397-1397) using random forest. These images were used as input for both classification and regression. In the initial map, written with Python, the three major classes of soil microforms, vegetation, and galls were identified and classified into five classes by classification (Figure 5).Changes of microlandforms with stochastic forest model in the period 1397-1397After classifying and creating a class on UAV images using random forest algorithm, changes in microlandforms were identified during the period 1397-1397. This algorithm with an average accuracy of 89%, determined that the changes of Afjeh microlandforms in the period 1397-1397 occurred in the range of vegetation and soil (Figure 6). So that the first rank of the most area changes in vegetation to soil was 66.64 and the second rank of soil to vegetation was 16.59% (Table 1, Figure 7); These changes were observed due to the greater dispersion of vegetation around the Jajroud River and the rock bed in the west of the study area. Unchanged areas in the bedrock of the river and soil cover and related microforms were observed in the central region of Afjeh.

In addition to field observations, the accuracy criterion was used to evaluate the evolution of microlandforms in the Afjeh region; The ratio of correctly classified pixels to total pixels was considered. The percentage given for the accuracy was obtained from this criterion. The study of the effect of this model on accuracy showed that the use of the features obtained from this network in the ability to distinguish different classes in UAV images is very useful. Thus creating a near-ideal model in the random forest algorithm. The results showed that the most changes of microlandforms in this model are related to the change of vegetation to soil (66.64%) and in the next rank is the change of soil to vegetation (16.59%). According to the obtained results, it was found that the flood in Afjeh region in 1397 has caused major changes in the region. As it has the greatest impact on the vegetation of the region, which has destroyed a large percentage of it. Vegetation-dependent microlandforms have also undergone major changes, so that the graph of its changes shows the highest level of turbulence compared to the more stable microlandforms of the Jajroud river bed. The diversion of the river is one of the changes observed in this period. In addition to the above, due to declining rainfall, recent droughts and floods; Sheetwash, ditches and galleys are more prevalent in the study area, indicating large changes in microlandforms and high erosion rates in the study area.