Assessing the Reliability of Landslides Susceptibility Models with Limited Data: Impact of Geomorphological Diversity and Technique Selection on Model Performance in Taounate Province, Northern Morocco

Previous research showed that the accuracy of landslide susceptibility maps (LSM) mainly depends on the landslide inventory used to train the algorithms. However, the preparation of the latter database is laborious and time-consuming, which is challenging especially in large study areas. For this reason, spatial extrapolation of LSM results obtained in small test zones can be valuable to researchers in limited data scenarios. In this paper, an extrapolation effort is conducted in two distinct geomorphological settings within the boundaries of Taounate province, Northern Morocco. In each setting, two test zones were considered for LSM analyses. Then the results were extrapolated on zone A (North of the study area), characterized by a heterogenous geomorphological setting, and the geomorphologically monotonous zone B located to the South. Our findings show that the extrapolated models yield acceptable to very good performances (0.65 < AUC < 0.85) depending on the geomorphological context, the number of the test zones selected and the technique used. By order of performance, which is defined in this paper by low sensitivity and high reliability, the techniques used can be ranked as follow, logistic regression, random forest, artificial neural networks, frequency ratio and lastly Shannon entropy. This is especially true in zone B where the monotonous landforms distribution is characterized by a lower sensitivity (less than 2%) compared to the heterogenous zone A. With respect to validation practices, our findings show that the validation strategy adopted by the landslides researchers significantly changes the accuracy assessment results (a 20% difference between expert derived and random approach). Therefore, testing data selection must be thoroughly and thoughtfully explained before proceeding to the validation of the LSMs.