Search

In January and February 2001, El Salvador was hit by two strong earthquakes that triggered thousands of landslides, causing 1,259 fatalities and extensive damage. The analysis of aerial and SPOT-4 satellite images taken a few days after the events allowed us to map 6,491 coesismic landslides, which occurred in 14 study areas extending for about 400 km2. Four different Multivariate Adaptive Regression Splines (MARS) models were produced by using different covariate sets and landslide inventories, the latter containing the slope failures triggered by an extreme rainfall event of November 2009 and those induced by the earthquakes of 2001. Moreover, two validation scenarios were employed, including the 25% and 95% of the mapped landslides, respectively. The results of our experiment revealed that: (i) MARS algorithm provides reliable predictions of coesismic landslides; (ii) models calibrated with rainfall-induced landslides predict with acceptable accuracy landslides caused by deep earthquakes and distributed over vast areas; (iii) the best accuracy is achieved by models trained with both preparatory and trigger variables; (iv) a small portion of the landslides produced by an earthquake can be used to calibrate MARS predictive models that help to identify slopes where yet unreported landslides may have occurred.

Landslide susceptibility assessment implemented by statistical methods relies on a basic concept according to which past and future landslides depends on the same causes of past failures. As a consequence, statistical inference can explore the relationships between past phenomena and geo-environmental variables to spatially recognize landslide-prone areas. Coherently, the quality and the prediction skill of the model and the relative prediction image heavily depend on the completeness of the landslide scenario exploited. However, landslide archives are frequently biased due to more or less limits we face in reconstructing reliable and complete landslide maps. In this sense, exploiting public available archives compiled from local administration is a goal in landslides susceptibility evaluation. In fact, usually, these types of archives are multitemporal and collect potentially or real destructive phenomena based on direct observation of failures. The use of these types of archives allows obtaining immediately available landslides inventories, saving time and resources from mapping.On the other hand, public landslide inventory may be wealthy close to urban areas while, usually, failures in the agricultural lands are sporadic or under-represented.In this study, starting from the available slide (56 cases) and flow (115 cases) inventories of the Torto River basin (420 km2, central-northern Sicily), which were prepared by the “Dipartimento Regionale dell’Autorità di Bacino del Distretto Idrografico Sicilia”, and a set of twelve geo-environmental predictors, two basin susceptibility models (for slides and flows, respectively) were prepared by applying Multivariate Adaptive Regression Splines (MARS). Good performance for the basin-scale models was then assessed according to cross-validation validation strategies. Then, in a randomly selected small sub-catchment (Sciara stream, 21.5 km2), the prediction images produced at a basin-scale were validated in recalling the landslides of two local systematic and multitemporal remotely recognized inventories. Despite the high performance of the basin-scale models, the results at the local scale showed a poor capacity of the models in detecting the two systematic archives, with a non-acceptable sensitivity (0.67 and 0.57 for slide and flow, respectively). The AUC (Area Under the Curve) values are non-acceptable (0.47 and 0.65 for slide and flow, respectively) also.In order to “boost” the basin-scale susceptibility models, by exploiting a score weighted random selection of 30% of the mapping units of the Sciara stream sub-catchment, a stable/unstable status concerning slide and flow phenomena of the selected cases was assigned by intersecting with the remotely sensed inventories.Two new basin-scale models were implemented by using these new “hybrid” archives. The two prediction images compared with the systematic inventories of the Sciara sub-catchment reveal an increment of the models’ performance, with high accuracy in predicting positive cases both for slide and local flow types. At the same time, persistent precision in detecting stable cases for local flow arises (~0.8), while a decrease in specificity due to an increment of False Positives suggests potentially new future activations for slide phenomena. However, the important increment of the AUC values (from 0.79 to 0.94 and from 0.8 to 0.94, for slide and local flow, respectively) testifies to a general improvement of the main models.

In this study, an inventory of storm-triggered debris flows performed in the area of the San Vicente volcano (El Salvador, CA) was used to calibrate predictive models and prepare a landslide susceptibility map. The storm event struck the area in November 2009 as the result of the simultaneous action of low-pressure system 96E and Hurricane Ida. Multivariate Adaptive Regression Splines (MARS) was employed to model the relationships between a set of environmental variables and the locations of the debris flows. Validation of the models was performed by splitting 100 random samples of event and non-event 10 m pixels into training and test subsets. The validation results revealed an excellent (area under the receiver operating characteristic (ROC) curve (AUC) = 0.80) and stable (AUC std. dev. = 0.01) ability of MARS to predict the locations of the debris flows which occurred in the study area. However, when using the Youden&rsquo, s index as probability threshold to discriminate between pixels predicted as positives and negatives, MARS exhibits a moderate ability to identify stable cells (specificity = 0.66). The final debris flow susceptibility map, which was prepared by averaging for each pixel the score of the 100 MARS repetitions, shows where future debris flows are more likely to occur, and thus may help in mitigating the risk associated with these landslides.

In the framework of the CARG (Geological and Geomorphological Mapping of Italy) project, landslides are also mapped as constituting one of the main surficial layers, masking the bedrock lithologies and related stratigraphic/tectonic contacts. As such, they are frequently mapped with a low resolution both in terms of spatial pattern and typology characterization. In particular, typical landslides affecting slopes in the Italian Apennines (slides and flows) are frequently grouped inside large polygons sometimes at a small catchment scale. However, the possibility to exploit such a reference landslide inventory for landslide susceptibility assessment is of great importance.In this test, the existing CARG landslide dataset for the “Visso” map (Marche, Italy) was split according to the movement typology by exploiting topography maps and orthophotos, thus producing rotational slides, earth flows, and complex landslides archives (198, 91, and 51 cases, respectively). Multivariate Adaptive Regression Splines (MARS)-based susceptibility models were following prepared by regressing each systematic landslide archive to a specific set of physical-environmental predictors, considered as determining for landslides activation. Furthermore, multicollinearity and variables importance analyses were carried out to verify their relevance and influence in landslide susceptibility assessment. Besides, a new type (LCL_SLU) of slope units, obtained by crossing classic hydrological partitioning with landform classification, was used as mapping units.The results show good AUC (Area under the ROC curve) for all models when prediction skill is evaluated, with values of 0.82, 0.77, 0.78 for rotational slides, earth flows, and complex landslides, respectively; the same AUC became outstanding when success skill is detected, with 0.91, 0.95, and 0.99 scores, respectively. Finally, for potential use in territorial planning, an integrated map was produced by adding up the single-landslide susceptibility scores and ranking the output on a classical 0-1 scale. The final map reaches an AUC value of 0.89, confirming the high performance of the models.The results of the test in the “Visso” map suggest as potentially very worth processing the landslide inventories already available from the CARG project to assess landslide susceptibility on a regional to national scale.

Landslide susceptibility can be evaluated by using different statistical approaches. Among these, the methods based on conditional analysis exploit the observed incidence of landslides into homogeneous statistical domains (corresponding to single classes of each geo-environmental variable or to multivariate Unique Condition Units) to estimate their landslide susceptibility. Thus, the results of these types of analysis can be heavily compromised by the completeness or representativeness of the adopted landslide archive. On the other hand, inference-based frequentist methods allow scoring landslide susceptibility by using limited samples of cases, provided the calibration samples are statistically representative of the whole population, assuming that the lacking cases are missing completely at random.This research aims to evaluate the effect of incomplete inventories in assessing landslide susceptibility, by using conditional analysis (Weight of Evidence, WoE; Frequency Ratio, FR) and inference-based (Binary Logistic Regression, BLR; Multivariate Adaptive Regression Splines, MARS) methods. In particular, we analysed the effects in terms of prediction skill of each of the four methods by reducing and randomly hiding the training calibration cases (and increasing the related validation cases).The study was conducted in the Imera Settentrionale river basin (Sicily, Italy), by exploiting two different landslide archives (5134 earth flow and 1608 rotational/translational slides) and a set of 10 physical-environmental predictors. Cutoff-dependent and -independent metrics (ROC-curve analysis and confusion matrixes) were used to estimate the performance of the models.As general assumptions, MARS and BLR modeling resulted as markedly more performing with moderately and asymptotically AUC improving up to 30-40% of the whole dataset, corresponding to the reaching of the relative optimal performance. A similar asymptotic AUC-increasing trend is described for WoE and FR, but with a lower performance. In particular, the optimal AUC values for rotational/translational slides range between 0.77 and 0.90, for BLR, 0.82 and 0.90, for MARS, 0.78 and 0.80, for FR, 0.76 and 0.78, for WoE. At the same time, a general lower model performance resulted for earth flows, with AUC values ranges of 0.69 and 0.75, for BLR; 0.75 and 0.79 for MARS; 0.67 and 0.70 for FR; 0.56 and 0.6, for WoE. Furthermore, differences in the selected predictors produced by the cases reduction were also explored through the analysis of the variable importance and the response curves.

To study on a regional basis, the relation between fluvial sediment delivery and coastal erosion, the historical record of coastline migration of Sicily was analyzed with respect to the estimated sediment delivery to the coast obtained from the spatially distributed sediment delivery WaTEM/SEDEM model. The latter was directly acquired from the ESDAC database as a 25 m pixel layers, being based on the combination between the RUSLE model and a transport capacity routing algorithm.At the same time, the coastline-evolution (accretion/retreatment) data for 1960/1994 and 1994/2012 intervals were processed. This dataset, provided by ISPRA (Italian Institute for Environmental Protection and Research), is made by vectorial polygons, corresponding to erosion or accretion areas obtained by the intersection between two coastlines. The dataset contains polygons related to the 1960-1994 and 1994-2012 periods.Once a common baseline was extracted from 2019 satellite images, 22 Physiographic Units (PU) were identified. The PU was defined based on geomorphologic criteria and by assuming a null net sediment budget (null sediment transport between two PU neighboring). Each coastal PU was connected to its contributing fluvial basins, also assigning the expected sediment delivery at the coastline.To perform the analysis, cross profiles along the coastline were generated and intersected with the polygons, calculating a response value, in terms of retreatment or accretion, to each of the cross-profile centroids. Finally, for each PU, the cumulated variations were computed.PUs with significant cumulative variations (more than 2 km) in at least one of the two epochs were identified and three different patterns were detected: accretion/retreatment, retreatment/accretion, and retreatment/retreatment. The response observed for the different PUs was then analyzed considering estimated sediment delivery, recognizing coherent (large sediment delivery = accretion) and incoherent (large sediment delivery = retreatment) behaviors, which have been interpreted as controlled by the history of soil/coastal erosion management practices.In particular, in spite of a very high expected sediment delivery, more than three-quarters of the Tyrrhenian coast resulted as affected by a marked retreat in 60-94 (same tens of meters) and a moderate accretion in 94-12, as the result of extensive coastal works which have been realized to mitigate coastal erosion.

The geodynamic context in which El Salvador is located, made of a convergent structure characterized by the interaction among six different plates, together with the lithological characteristics of the outcropping rocks and soils (mainly corresponding to deeply weathered acid pyroclastites, basic effusive rocks and volcanic ashes), are responsible for the very high seismically-induced landslide susceptibility of the country. These predisposing factors were decisive on the occurrence of thousands of seismically-induced landslides caused by two huge earthquakes on 13th January and 13th February 2001, which triggered thousands of landslides in the country. In particular, the February event (6.6M, onshore and intraplate at a depth of 10 km) triggered 5,371 landslides in an area of around 300km2. These gravitational phenomena took the form of debris slides, earth slides and debris flows and affected several inhabited areas damaging infrastructures and crops and causing, respectively 844 and 315 fatalities.Thanks to aerial photos taken soon after the days following both the two earthquakes and made available by the CNR (Centro Nacional de Registros - Instituto Geográfico y del Catastro Nacional), associated landslide maps have been prepared, where each phenomenon is represented by a landslide polygon and its LIP (Landslide Identification Point), located in the crown of the landslide. In particular, static landslide susceptibility models were prepared for the Ilopango (1594 landslides in an area of around 40km2) and the San Vicente (1602 landslides in an area of around 108 km2) sectors, by regressing the spatial distribution of the 13th February seismically-induced landslides on a set of explanatory variables obtained by a geologic map and a 10m pixel DTM (Digital Terrain Model). At the same time, shaking-dependent models were prepared by including also PGA (Peak Ground Acceleration) and the epicentral distance (ED) among the predictors.For both the two areas a marked increase of performance was observed (AUC from 0.70 to 0.75, for Ilopango, from 0.73 to 0.77, for San Vicente) from the static to the shaking-dependent models, highlighting the role of the seismic acceleration in the triggering of the landslides both in activating the susceptible sites and in lowering the score threshold for slope failures occurrences. Besides, for the Ilopango sector, a rainfall-induced susceptibility model was also prepared, exploiting a landslide inventory available for the 2009 IDA/12E storm events. The obtained score was then combined with PGA and ED to predict the spatial distribution of the seismically induced landslides, obtaining a higher performance than the relative basic model (AUC = 0.75).The results obtained from the research demonstrate suggest the possibility to couple the susceptibility scores obtained from static modelling to the expected mechanical shaking for the seismically-induced susceptibility assessment.The whole modelling was carried out by applying MARS (Multivariate Adaptive Regression Splines) analysis through RStudio and SAGA GIS freeware software.

Optimally sizing grid cells is a relevant research issue in landslide susceptibility evaluation. In fact, the size of the adopted mapping units influences several aspects spanning from statistical (the number of positive/negative cases and prevalence and resolution/precision trade-off) and purely geomorphological (the representativeness of the mapping units and the diagnostic areas) to cartographic (the suitability of the obtained prediction images for the final users) topics. In this paper, the results of landslide susceptibility modelling in a 343 km2 catchment for three different types of landslides (rotational/translational slides, slope flows and local flows) using different pixel-size mapping units (5, 8, 10, 16 and 32 m) are compared and discussed. The obtained results show that the higher-resolution model (5 m) did not produce the best performance for any of the landslide typologies. The model with 8 m sized pixels displayed the optimal threshold size for slides and slope flows. In contrast, for local flows, an increasing trend of model prediction accuracy was reached with 32 m pixels, which was a higher value than that presented using 8 m pixels. The variable importance analysis demonstrated that the better performance of the 8 m cells was due to their effectiveness in capturing morphological conditions which favour slope instability (profile curvature and middle and high ridges).

Sediments accumulated in check dams are a valuable measure to estimate soil erosion rates. Here, geographic information systems (GIS) and three machine learning techniques (MARS-multivariate adaptive regression splines, RF-random forest and SVM-support vector machine) were used, for the first time, to predict sediment deposition rate (SR) in check-dams located in six watersheds in SW Spain. There, 160 dry-stone check dams (~ 77.8 check-dams km−2), accumulated sediments during a period that varied from 11 to 23years. The SR was estimated in former research using a topographical method and a high-resolution Digital Elevation Model (DEM) (average of 0.14 m3ha−1 year−1). Nine environmental-topographic parameters were calculated and employed as predictors of the SR. The ability of MARS, RF and SVM was evaluated by using a five-fold cross-validation, considering the entire area (ALL), the check dams on the hillslope (HILL) and the valley-bottoms (VALLEY), as well as the three catchments (B, C and D) with the highest number of check dams. The accuracy of the models was assessed by the relative root mean square error (RRMSE) and the mean absolute error (MAE). The results revealed that RF and SVM are able to predict SR with higher and more stable accuracy than MARS. This is evident for the datasets ALL, VALLEY and D, where errors of prediction exhibited by MARS were from 44 to 77% (RRMSE) and from 37 to 62% (MAE) higher than those achieved by RF and SVM, but it also held for the datasets HILL and B where the difference of RRMSE and MAE was 7–10% and 12–17%, respectively.

In the studies of landslide susceptibility assessment, which have been developed in recent years, statistical methods have increasingly been applied. Among all, the BLR (Binary Logistic Regression) certainly finds a more extensive application while MARS (Multivariate Adaptive Regression Splines), despite the good performance and the innovation of the strategies of analysis, only recently began to be employed as a statistical tool for predicting landslide occurrence. The purpose of this research was to evaluate the predictive performance and identify possible drawbacks of the two statistical techniques mentioned above, focusing in particular on the prediction of debris flows. To this aim, an inventory of debris flows triggered by the passage of the hurricane IDA and the low-pressure system associated with it 96E, on 7 th and 8 th November 2009, in an area of about 26 km 2 close to the Caldera Ilopango, El Salvador (CA), was employed. Two validation strategies have been applied to both statistical techniques, thus obtaining four models – BLR (I), MARS (I), BLR (II) and MARS (II) – to be compared in pairs. Model performance was assessed in terms of AUC (area under the receiver operating characteristic (ROC) curve), Sensitivity, Specificity, Positive Prediction Value and Negative Prediction Value. Moreover, to evaluate the robustness of the modelling procedure, 50 replicates were created for each model and standard deviation was calculated for each of them. The results show that both techniques allow for obtaining good or excellent performances so that it is not possible to define one of the two techniques as absolutely better. However, the validation procedure reveals slightly better performance of the MARS models, with greater sensitivity and greater discrimination among True Negatives (TNs).

The main topic of this research was to evaluate the effect in the performance of stochastic landslide susceptibility models, produced by differences between the triggering events of the calibration and validation datasets. In the Caldera Ilopango area (El Salvador), MARS (multivariate adaptive regression splines)-based susceptibility modeling was applied using a set of physical–environmental predictors and two remotely recognized landslide inventories: one dated at 2003 (1503 landslides), which was the result of a normal rainfall season, and one which was produced by the combined effect of the Ida hurricane and the 96E tropical depression in 2009 (2237 landslides). Both the event inventories included shallow debris—flow or slide landslides, which involved the weathered mantle of the pyroclastic rocks that largely outcrop in the study area. To this aim, different model building and validation strategies were applied (self-validation, forward and backward chrono-validations), and their performances evaluated both through cutoff-dependent and -independent metrics. All of the tested models produced largely acceptable AUC (area under curve) values, albeit a loss in the predictive performance from self-validation to chrono-validation was observed. Besides, in terms of positive/negative predictions, some critical differences arose: using the 2009 extreme landslide inventory for calibration resulted in higher sensitivity but lower specificity; conversely, using the 2003 normal trigger landslide calibration inventory led to higher specificity but lower sensitivity, with relevant increase in type-II errors. These results suggest the need for investigating the extent of such effects, taking multitrigger intensities inventories as a standard procedure for susceptibility assessment in areas where extreme events potentially occur.