**5. Discussion**

From the obtained results (Table 2), it is evident that the choice of algorithm and sampling strategies can affect the prediction performance of a landslide susceptibility map significantly. The effect of data splitting is crucial for only RF, KNN, and SVM algorithms while using the point data for sampling. The landslide susceptibility maps and H-index plots provide more insights into the effects of different sampling strategies in the performance of different algorithms. From the H-index maps and AUC values, it is evident that the sampling strategy is least effective in the case of NB and most effective in the case of RF.

Figure 12 shows the H-index plots prepared to understand the mutual agreemen<sup>t</sup> between different algorithms using the same sampling strategy. It can be observed that, in the case of low susceptible area, the different algorithms are in good agreemen<sup>t</sup> with each other, and the LR algorithm classifies the least area in the very low category, which is 51.30% of the total area. While using point data, all algorithms agree in the classification of 47.56% of the total area and all algorithms differ in the case of 0.17% of the total area.

**Figure 12.** H-index maps plotted using all five algorithms with: (**a**) point data, and (**b**) polygon data.

The percentage distribution of each value of H-index is provided in Table 3 below. While using polygon data, the mutual agreemen<sup>t</sup> between algorithms is improved, with perfect agreemen<sup>t</sup> in 58.06% of the total area. In no pixels, the classification of all algorithms is entirely different and at least two algorithms agree with the predicted classification. As can be observed from Figure 12b and Table 3, there are no pixels with a H-index value of 0.70 when polygon data is used.

For NB and LR algorithms, the performance is reduced when a greater number of data points in the polygon dataset is used. This is a result of increased correlation between the LCFs with more data points, which violates the basic assumption of independent variables in both the cases. The use of linear fitting function in the case of LR also results in a slight decrease in the accuracy and AUC values with the increased number of data points. However, the advantage of using these algorithms is the reduced computational time involved, as they do not require any hyper parameter tuning.


**Table 3.** Percentage distribution of H-index values in the total area, using different sampling strategies: comparison between all algorithms.

In the case of KNN, SVM, and RF, the ratio of the train to test dataset can also result in a performance variation while using point data. The performance of these algorithms is significantly increased with the use of polygon data. The improvement in performance can be attributed to the improved size of data used for training the model. All three models demand a long time for the fine-tuning process. The models are highly sensitive to the parameters, train to test ratio, and the size of the dataset [5]. All the three models are widely used for LSM and, hence, if computational facilities are available, the train to test ratio should also be varied to produce the best results from these algorithms.

Even though the performance is comparable with KNN and RF, a higher number of landslides in the very low category make the landslide susceptibility maps made using SVM unsuitable for practical applications. This is an important aspect to be considered. From Figure 11, it is evident that the model using polygon data with an AUC of 0.963 is classifying 13% of the landslides in the very low susceptible zone. This is visible in the landslide susceptibility maps in Figure 11b. The performance can be further improved by using different data sampling approaches and ensemble algorithms and neural networks. In the case of RF, even though the results are statistically better from a practical perspective, the very high, high, and medium classes are bounded by the polygon data used for training and the model is too optimistic, which does not leave room for possible landslides in the surrounding areas in the future. The same issue is observed with the landslide susceptibility map prepared with the KNN algorithm using polygon data. Even though these three algorithms (KNN, SVM, and RF) are having the highest statistical attributes, they cannot be considered to be the best suited for the landslide susceptibility map, due to the limited part of the study area classified into very high, high, and medium classes. The landslide susceptibility map must be conservative, which considers the possible occurrence of landslides in areas other than the ones used for training and testing, and, at the same time, should not classify the safe zones as landslide-susceptible regions. The landslide susceptibility map produced using the RF algorithm with point data is an optimum solution with good statistical performance (AUC = 0.952 and accuracy = 88.12%) and practical applications. It classifies 7.87% of the total area into the very high category and 9.79%, 7.09%, 15.17%, and 60.08% into the high, medium, low, and very low categories, while the best performing model is developed using RF with polygon dataset, with an accuracy of 97.30% and an AUC of 0.993.

From the results, it can be inferred that both the choice of algorithm and sampling strategy can influence the prediction performance of LSM, but the choice of the landslide susceptibility map should not be based on the statistical performance only.
