Earthquakes are natural events that cause great damage to nature, buildings, engineering structures, and cause human death (Gorum and Carranza 2015). Moderate and severe magnitude earthquakes trigger thousands of landslides, especially in rugged and high-slope mountainous regions (Gorum et al., 2013; Tanyaş et al., 2017). For example, in the 2008 Wenchuan, China Mw 7.9 earthquake, there were about 200,000 landslides (Tian et al., 2019), about 5,000 landslides in the January 12, 2010 Mw 7.0 Haiti earthquake (Gorum et al., 2013), more than 2,000 landslides in the 2015 Gorkha, Nepal Mw 7.8 earthquake, and about 6,000 landslides in the 2018 Hokkaido Iburu Mw 6.7 earthquake (Yamagishi and Yamazaki 2018). Landslides that appear as a secondary effect of earthquakes cause human deaths and economic losses. Fatal landslides occurred in 76 of the 196 earthquakes between 1811 and 2016. These fatalities correspond to 17.7% (213,913 people) of deaths caused by earthquakes (Jessee et al., 2020). Earthquake-triggered landslides accounted for 5.2% ($ 170 billion) of economic damage from earthquakes between 1900 and 2016 (Daniell et al., 2017). Therefore, rapid mapping of landslides is important for rapid response to disaster areas and crisis management.

Active and passive remote sensing systems offer great advantages in rapid landslide mapping (Aimaiti et al., 2019). With remote sensing systems, Optical images (Zhao et al., 2017; Shao et al., 2019), synthetic aperture radar (SAR) (Aimaiti et al., 2019; Adriano et al., 2020), LIDAR systems (Liu et al., 2019), Unmanned Aerial Vehicle (UAV) systems (Comert et al., 2019) or synthesis of optical and SAR images (Shirvani et al., 2019) can be used in the landslide mapping. The most common method preferred in mapping landslides is visual image interpretation (Guzzetti et al., 2012; Rosi et al., 2018). Although the image interpretation provides high accuracy, it is a slow method that is not suitable at the time of rapid intervention since the landslides are mapped by the expert with image interpretation and manual digitization (Guzzetti et al., 2012). Therefore, automatic landslide mapping is an important factor for post-disaster crisis management. Image classification approaches (Vamsee et al., 2018) and change detection methods (Lu et al., 2019) are used in automatic landslide mapping. Since change detection methods require pre-event images, they present a disadvantage in mapping the landslide events triggered by the earthquake. However, it is a useful method in case of the pre-event image available (Yang and Chen 2010; Lu et al., 2019). Classification methods are generally carried out on post-event images (Stumpf and Kerle 2011; Dou et al., 2015). The disadvantage of this method in making rapid decisions is that it takes time to generate the training dataset. However, when the classification model is developed over the appropriate training dataset, the mapping process is shortened (Danneels et al., 2007; Stumpf and Kerle 2011; Mondini et al., 2011).

Image classification is used in landslide mapping in two approaches, pixel-based image analysis (PBIA) (Danneels et al., 2007) and object-based image analysis (OBIA) (Martha et al., 2010; Stumpf and Kerle 2011; Hölbling et al., 2015; Shirvani et al., 2019). After the landslide events, different sizes of landslides can occur in the relevant area. High spatial resolution images offer great advantages to detect landslides of different sizes. OBIA has become more popular in landslide mapping in the last decade since misclassified pixels have emerged from the landslide mapping from high spatial resolution images with PBIA (Guzzetti et al., 2012; Zhong et al., 2020). OBIA is applied in two different ways in landslide mapping: rule-based (Martha et al., 2010; Hölbling et al., 2012; Comert et al., 2019) and supervised mapping (Stumpf and Kerle 2011; Shirvani et al., 2019). In the landslide mapping process, landslides should be separated from different land use classes. A large number of object features must be used for this process. Rule-based mapping is a slow method for complex areas. since the object features and threshold values that separate the classes must be determined by an expert (Stumpf and Kerle, 2011). In supervised mapping performed with a selected algorithm such as random forest, the features used to separate the classes are automatically determined, and rapid results are produced by the mapping process made on the training data to be produced in the appropriate size and number (Stumpf and Kerle 2011; Shirvani et al., 2019; Shirvani, 2020). Therefore, supervised mapping process will give rapid results in rapid intervention in crisis management.

In this study, the success of the size of the model area selected for training data, and the dataset type to be used in classification in the rapid detection of earthquake-triggered landslides were investigated. In particular, the following questions have been addressed:

• How enlarging the model area selected for training data affect earthquake-triggered landslide mapping accuracy?

For these purposes, the Hokkaido Iburu earthquake zone in Japan was chosen as the study area. PlanetScope images and ALOS-PALSAR Digital Elevation Model (DEM) were used as analysis data. OBIA was preferred as the mapping approach and the Random Forest (RF) algorithm was used as the classification method in the OBIA. RF algorithm, which is a fast machine learning classifier, was preferred in this study because it gives successful results in many areas such as landslide susceptibility mapping (Dou et al., 2019; Shirvani 2020), landslide mapping (Stumpf and Kerle 2011; Chen et al., 2017; Shirvani et al., 2019; Maxwell et al., 2020), burnt area mapping (Ramo and Chuvieco 2017).

The methodology applied within the scope of the study is shown in Figure 2. The landslide mapping process was carried out using the OBIA approach. OBIA and PBIA are two different approaches used in image classification. Spectral values of pixels are used when classifying images with PBIA, while spectral, geometric, textural, and spatial values of image objects are used for classification with OBIA. OBIA was preferred in this study because it reduces the problem of misclassification error in high spatial resolution images includes different image object metrics in addition to spectral features and reduces the data size (Zhong et al., 2020).

In the segmentation stage, different combinations of post-event image bands and layers produced for image difference were tried to identify appropriate input bands to create image objects representing landslides. As a result of the experiments, it was seen that the landslide objects were best obtained from the post-event R, G, B, NIR bands. Therefore, these layers were used as input bands in the segmentation process for all data sets. Different scale parameters (from 10 to 150 by 10), shape, compactness, and layer weights were tested in determining the optimum parameters for MRS. As a result of trial and error, suitable parameters were determined as a scale: 110, shape: 0.3, compactness: 0.7, layer weights as B = G = NIR = 1, R = 2. These values have been applied to be the same for all data sets. These values were used to create image objects for all data sets. Landslides in the image objects obtained result of the segmentation process consists of more than one image object. When the scale parameter was set to more than 110, it has been observed that especially small landslides mix with other land use classes. Image objects created by segmentation for model and test areas were labeled as Landslide (LS) and Non-Landslide (NLS) using pre-event and post-event images. The data sets created for the RF classifier and the numbers of image objects belonging to LS and NLS included in them are shown in Table 4.

The RF classifier development process was carried out using WEKA (Waikato Environment for Knowledge Analysis) data mining software. The process of determining the optimum number of trees and the minimum variable was carried out separately for each model area. The number of random trees was applied as different values from 100 to 1500. The number of m variables was applied to the model data sets as 6, 7, 8, 9 values. As a result of the experiments, the m value was determined as 8 for all data sets. The number of trees was defined as 600 for model areas 2 and model area 3 in the change-based dataset. The number of trees in model area 4 was determined as 400 for both the change-based dataset and the post-event dataset. The number of trees was set as 500 for the remaining classifiers. The success of the RF classifiers created for model areas according to the determined parameters on the test area is shown in Table 5. The overall accuracy success of RF classifiers on the test site for separating LS and NLS areas has been achieved by over 95% with all classifiers in both datasets (Table 5).

When the accuracy results were examined (Table 5), it was seen that TP Rate increases in the post-event dataset when the model area expands. With the classifier developed for Model Area 1, 92% of the image objects belonging to the LS in the test area were successfully mapped. This ratio was obtained with 3% more accuracy in the largest dataset. When the number of training data is increased by expanding the model area, the FP rate, which defines the mapping rate of NLS image objects as LS, slightly increased. This change had reduced the precision rate of classifiers. When the success of the developed classifiers for landslide mapping was evaluated, it is observed that very close F-measure values were obtained, but a little more success was achieved in the largest dataset.

When the accuracies for the change-based datasets are investigated, the TP ratio of LS image objects was obtained approximately 90% for the smallest dataset. This ratio decreased by 1 and 2% in model area 2 and model area 3, respectively. In the largest dataset, very close results were obtained with model area 1. When the size of the model dataset was increased, it was seen that the sensitivity of the FP ratio increased and the extraction of NLS objects as LS decreased. This condition has increased the precision ratio and the F-measure value of the developed classifiers.

Comparing the results obtained for the change-based dataset and post-event dataset, landslides were obtained with a higher TP Rate in all model areas where RF classifiers were developed. When the FP Rate values were compared, it was seen that when the data size was increased, the classification of NLS image objects as LS with the change-based data set decreased, and the developed classifiers map the NLS areas with higher accuracy. When the overall success of the classifiers for two datasets was evaluated according to F-measure, it was indicated that landslide mapping with post-event dataset gave higher accuracy values.

In this study, 40 different features were used for the change-based dataset and 43 different features for the post-event dataset. When the feature importance of the variables was investigated, it was observed that the same variables were effective in all model areas. δNDVI in change-based RF classifiers and slope and NDVI in post-event RF classifiers were the most effective features. In Figure 3, the top ten features in the classifier created for the model area 5 based on change-based datasets were shown according to the mean decrease of accuracy and mean decrease of Gini index. When Figure 3 is examined, it is seen that the most important parameters in classification according to mean decrease accuracy were mean values of δNDVI, slope, PCA3, and ICA2. The most important variables that affect the purity of the classifier when extracted from the dataset were mean values of δNDVI, ICA3, PCA3, and PCA4. In Figure 4, the top ten features obtained for Model Area 5 with the post-event dataset were given. The most important features for the developed classifier were obtained as mean values of the slope, NDVI, PCA4, PCA3, and Maximum Difference. The most important parameters that increase the purity of the model were NDVI, PCA2, NDWI, Max Difference, and Red (Figure 4). When the importance of properties for different datasets is evaluated, it is seen that spectral properties and spatial properties are more effective in mapping landslides according to textural properties.

In this study, the influence of model development area size and dataset types in landslide mapping during the post-earthquake crisis management phase was investigated. In this context, ten landslide mapping models have been developed with OBIA-RF classifier using two different datasets, namely change-based and post-event dataset, at five model development areas with different sizes. The success of the developed classifier models has been investigated on a determined test area. When the obtained results were examined, although landslide mapping results were obtained with high accuracy from all developed models, higher accuracy values were obtained in classifier models using post-event datasets (Table 5). These results show that while pre-event and post-event images are useful to create or update the landslide inventory maps (Yang and Chen, 2010; Đurić et al., 2017), post-event images without the need for any archive images in crisis management are sufficient in detecting the locations where the dense landslide events occurred. In the comparison of dataset size in this study, it was observed that increasing the dataset size slightly increased the mapping accuracy (Table 5). Therefore, large model areas are not required in the post-disaster decision-making phase. The model development process, which is carried out on a small area that best represents the area to be mapped, is useful for rapid decision-making.

The RF algorithm for landslide mapping was applied to datasets produced by using PlanetScope images and ALOS-PALSAR DEM data in the study. The RF algorithm, which gives successful results in extracting the desired feature in datasets with large sizes and many variables in different application areas (Stumpf and Kerle, 2011; Ramo and Chuvieco, 2017; Shirvani et al., 2019; Maxwell et al., 2020; Shirvani, 2020), has also produced high accuracy results from the datasets used (Table 5). With this research, it has been demonstrated that PlanetScope satellites, which offer daily high spatial resolution images, provide successful results in the detection of earthquake-triggered landslide areas with high accuracy. To produce the slope map of the study area, 30 m resolution ALOS-PALSAR DEM data were used as elevation data. Although this DEM is sufficient for this application in terms of presenting the general slope characteristics of topographic surfaces where landslides occur in the study area, it is insufficient in terms of producing induvial landslide inventory mapping. During the segmentation phase, landslides were obtained as multiple image objects. By merging these image objects, holistic landslides can be obtained (Figure 5). However, these landslides were not individual. The main reason for this is the high reflection of fresh landslides in satellite bands. In the segmentation stage, adjacent landslides with the same spectral value fall into the same image object. For these landslides to be used in susceptibility and hazard maps, the necessary ones should be converted into individual landslides. For this process, post-event high-resolution DEM data of the area is required (Marc and Hovius 2015).

While determining the object features in the datasets, firstly, the visual evaluation of the image object features to be included in the datasets on the image objects created as a result of the segmentation stage was made. Since the landslide areas consist of more than one image object, geometric features such as length/width, size, shape were not added in the datasets because there was no distinctive finding was observed on these features. Variable importance analysis was performed for the image object features used in datasets. The ten most important features for the datasets were shown in Figures 3, 4. NDVI, which is an effective feature in separating green vegetation areas from landslides, is a frequently preferred feature in landslide mapping with change detection (Mondini et al., 2011; Hölbling et al., 2015; Lu et al., 2019) and classification method (Martha et al., 2010; Shirvani et al., 2019). Since the application areas were chosen in this study were covered with dense forest areas and agricultural land, it has been revealed that NDVI and δNDVI were important features in both the change-based dataset and the post-event dataset (Figures 3, 4). The slope value is especially in the separation of low-slope areas such as roads and bare soil spectrally similar to landslides from landslides (Comert et al., 2019). The slope feature was seen to be an important feature that increases the accuracy of landslide mapping for both the change-based dataset and post-event dataset in this study (Figures 3, 4).

PCA and ICA, which are non-parametric feature extraction methods, are the preferred approaches in landslide mapping with change detection (Mondini et al., 2011; Lu et al., 2019). PCA4, ICA2, ICA3, and ICA4 components produced from pre-event and post-event NIR and Red bands gave successful results (Lu et al., 2019). In this study, PCA1_change, PCA3_change, and PCA4_change, which were among the PCA components produced for the change-based dataset, were found to be important features that increase the accuracy of the classifiers and ensure the purity of the decision trees used in landslide mapping (Figure 3). ICA3, one of the ICA components produced for the change-based dataset, has been seen to be the second most important component that provides the homogeneity of decision trees used in the separation of LS and NLS areas (Figure 3). Also, ICA2 was the fourth feature that positively affects classifier accuracy (Figure 3). In the study, four PCA components were generated from post-event image bands and included in the post-event dataset. Among these features, PCA2_post, PCA3_post, and PCA4_post were found to be important features that increase classification accuracy and provide decision tree purity (Figure 4). According to the variable important results, it shows that the non-parametric components obtained by PCA analysis can be used in landslide mapping studies based on the post-event images.

When examining other features obtained by variable important analysis, it was seen that the Max. Diff. the feature was important in both classifier accuracy and model homogeneity for the post-event dataset. Also, Red, Green Blue bands were obtained as important properties in providing model purity for the post-event dataset. CDI is used in landslide mapping with change detection from multi-temporal images (Li et al., 2016; Lu et al., 2019). In this study, it was seen that for the change-based data set, it was an important parameter in terms of classifiers model homogeneity, although not as much as, δNDVI, PCA, and ICA components.

In this study, the success of different datasets and different data sizes produced from remote sensing images in rapid earthquake-triggered landslide mapping were investigated. In this context, change-based and post-event datasets were created for five model areas and one test area. When RF classifiers developed on model areas were applied to the test area, high accuracies are obtained for all model areas and datasets. However, slightly higher accuracy results have been achieved with post-event datasets compared to change-based dates. Similarly, slightly higher accuracy was obtained when the model area size was increased. When these results are evaluated, in rapid detection of the location where earthquake-triggered landslides occur: 1) models to be used in mapping landslides can be developed over a small area that best represents the area to be mapped, 2) in the landslide mapping process, landslides can be successfully detected from the post-event images without the archive image of the relevant area. In the study, the importance of image object features used in datasets in the mapping process was investigated. When the importance levels of the object features are examined, it has been observed that the highly important features are similar to the features used in the literature. Also, PCA components generated from post-event images are included in the post-event data. Since these components positively affect the classification accuracies, they can be included in the post-event datasets in future studies. The approach presented in this study is an appropriate method for detecting large landslide areas that occur after a natural disaster and to respond rapidly. The presented method needs improvement to create landslide inventory maps. In future studies, both crisis management and inventory mapping research will be conducted from platforms that offer higher spatial resolution images and DEMs such as Unmanned Aerial Vehicle.