Introduction

Rainfall-induced mass movements are one of the most frequent causes of natural disasters worldwide, particularly in tropical and mountainous countries of the circum-pacific region (Naranjo Bedoya et al. 2019). The variability in rainfall (distribution, amount, frequency, intensity, and duration) is by far the main trigger and the most common cause of mass movements in most areas of the world (Caine 1980; Guzzetti et al. 2008; Aristizábal and García-Aristizábal 2020; Gariano and Guzzetti 2022). Rainfall can influence slope stability indirectly, through its effect on pore water conditions in the slope material, and it is therefore crucial to understand the changes in rainfall patterns that may drive the damage from these hazards (Caine 1980; Emberson et al. 2021) causing individual landslides with localized impacts or large-scale extreme rainfall events that can induce numerous landslides widely spread over the landscape (Mtibaa and Tsunetaka 2023).

Globally, landslides triggered by rainfall are responsible for 89.6% of all landslide fatalities, and about 90% of deaths from landslides have been the result of rainfall-triggering events (Petley 2008). According to the World Health Organization (WHO), from 1998 to 2017, landslides impacted about 4.8 million individuals, resulting in over 18,000 global fatalities (Ortiz-Giraldo et al. 2023). In 2020, the United National Office for Disaster Risk Reduction (UNDRR) noted an average annual economic loss of US$ 151.6 billion. The EM-DAT database from the Centre for Research on the Epidemiology of Disasters (CRED) (https://www.emdat.be/database) recorded a comparable number of events in 2022 as in the preceding decade. The number of deaths and affected people (403 and 100,000) were nearly half of the figures for the 2002–2021 period (838 and 200,000), respectively. Landslides are common phenomena in mountainous tropical regions where rainfall intensity and duration can trigger shallow landslide (Cullen et al. 2022), and where these regions are experiencing increasingly unfavorable conditions caused by unplanned land use and the increased frequency of extreme rainfall events. More than 80% of fatal landslides occur in the tropics (Ozturk et al. 2022) and worldwide patterns on the loss of life caused by landslides clearly demonstrates that landslides pose a significant hazard in these regions (Vega et al. 2023; Vega and Hidalgo 2023). Therefore, the case of Colombia is no more encouraging, 88% of the occurred natural disasters had hydrometeorological origin, where mass movements account for 15%, ranking second after floods (World Bank 2012). The Mass Movements Information System (SIMMA, in Spanish) of the Colombian Geological Service (SGC, in Spanish), between 1900 and 2022, registered more than 32,000 landslides, where the rainfall showed a dominant influence on the occurrence with 92% of events (Cerón et al. 2022), with a greater influence of rainfall duration instead of intensity (Aristizabal et al. 2022), resulting in more than 31,198 fatalities and economic losses higher than US$ 654 million.

Global warming is changing rainfall and the hydrological cycle, especially the extremes (Trenberth 2011) and due to their role in accelerating extreme weather events, its anomalies, defined as deviations of the amount of rainfall from the long-term averages (Zaveri et al. 2020) falling in a given region over a specified period, can have significant implications because the knowledge of the spatial distribution and rainfall annual cycle is important for a country from a hydrological, environmental, social, and economic point of view. Likewise, El Niño–Southern Oscillation (ENSO) is causing year-to-year climatic variations on Earth (Suhaila 2021). In several regions of South America, there are significant impacts on seasonal and monthly rainfall amounts during the different phases of the ENSO (Grimm and Tedeschi 2009). For instance, Colombia is highly influenced by climate variability due to its tropical condition and its location in the Andes. The complex topography induces spatiotemporal hydroclimate variability at different scales and atmospheric circulations, and enhanced deep convection is induced as a fundamental mechanism in the formation of tropical rainfall, controlled mainly by both phases, warm and cold, of ENSO, which has been associated with greater landslide occurrence around the world (Moreiras 2005). In fact, under the influence of La Niña inducing severe rainy seasons, the largest period of fatalities during the 2010–2011 La Niña event, identified by Sepúlveda and Petley (2015), affected ~9% of the total population and caused economic losses of approximately US$ 7.8 billion (Hoyos et al. 2013).

In a literature review, few studies have been conducted in Colombia for the analysis of the relationship between rainfall, landslide occurrence, and the influence of the ENSO climate phenomenon. Moreno et al. (2006) analyzed the relationship between landslide occurrences, rainfall, and the influence of ENSO phases in the department of Antioquia at different temporal scales, establishing that La Niña, on an interannual scale, and the two wetter periods over the Andean zone, on an intra-annual scale, favored an increase in the number of landslides. Ramos-Cañón et al. (2016) analyzed time series of mass movements in Bogotá, and they observed a direct relationship with the Oceanic Niño Index (ONI) of the La Niña phase, as well as a lagged correlation between rainfall and the occurrence of mass movements, where most mass movements were triggered weeks after high rainfall values. More recently, Naranjo Bedoya et al. (2019) presented the spatial and temporal rainfall variability and their relationship with historical landslides events between 1985 and 2016, under the influence of ENSO in the nine sub-regions of the department of Antioquia, concluding that the occurrence of landslides exhibited bimodal behavior according to temporal rainfall variability, and the strong effect of the cold phase of the ENSO (La Niña) had a strong influence on them. Aristizábal and García-Aristizábal (2020) carried out a temporal analysis (1950–2018), which included the correlation between rainfall and landslides in the Aburrá Valley (department of Antioquia), observing the dependence of landslides on the type of rainfall cycle, with differences in the rainfall-slides correlations for El Niño, La Niña, and Neutral years, but it was unclear whether this same trend was observed throughout the Andean region. Gómez et al. (2021) analyzed the distribution of landslide occurrences with associated deaths in Colombia between 1921 and 2019, finding that rainfall was the most common cause, which is consistent with other studies that confirm the close relationship between rainfall and landslides. Aristizábal and Sánchez (2020) presented a detailed analysis of the spatial and temporal trends of landslide occurrence in Colombia from 1900 to 2018 where rainfall was the most common triggering factor with 92% of the landslides registered in the period analyzed, but the influence of ENSO on the landslides was not evaluated. Gómez et al. (2023b) presented an analysis of four widely known global databases (1903–2020), which contain relevant landslide information for different regions of the world showing the American continent as one with the highest number of landslides and associated fatalities, showing a bimodal annual temporal pattern. Cullen et al. (2022) introduced a satellite-based landslide rainfall threshold for Colombia by studying 4 years (2016–2019) of rainfall measurements. A bimodal annual pattern of landslide occurrence was observed in the data, with two maximum peaks in May and October–November, reflecting the strong influence of rainfall on landslide occurrence. The most comprehensive approach was conducted by Garcia-Delgado et al. (2022), providing a detailed assessment of fatal landslides in Colombia for a long time interval (1912–2020). They identified that the landslides presented a bimodal temporal distribution that approximately coincided with the distribution of rainfall in the main cities of the country. Likewise, they found that in general, in Strong La Niña years, the number of fatalities is greater than in El Niño periods. However, the spatial variability of rainfall was not evaluated.

To enhance the comprehension of the intricate spatiotemporal connections and dynamics between rainfall and landslides in tropical mountain regions, and to effectively evaluate landscape changes in support of hazard prediction efforts, a thorough analysis of their interlinkage is essential across diverse spatial and temporal scales. Consequently, this study conducts a comprehensive space-time and frequency analysis to elucidate the relationship among landslides, rainfall variability, and ENSO in the Tropical Andean Mountain region of Colombia. Leveraging a 42-year dataset (1981–2022) encompassing rainfall and Multivariate ENSO Index v2.0 time series, alongside records from two landslide databases, this research offers valuable insights. Figure 1 visually presents the spatial and temporal distribution of the analyzed landslide events categorized over 6-year intervals.

Fig. 1
figure 1

a Geographical context of Colombia. The elevation is highlighted in Andean region. b Main departments within Andean region. c Space-time distribution of the analyzed landslides records categorized by 6-year period. d Color bar charts with the number of landslides in the same 6-year periods and black dotted line with the cumulative landslide events

Full size image

Study area

Colombia is located in the northwest corner of South America (Fig. 1a). It is crossed by the Andes Mountain range, giving rise to several climates and ecosystems. Colombia has been divided into five natural regions defined by the Geographic Institute Agustín Codazzi (IGAC, in Spanish), Caribbean, Pacific, Andean (Fig. 1b), Orinoco, and Amazon.

Geological and geomorphological context

The geomorphological configuration of the Colombian Andean ranges is the result of a convergence zone by subduction processes in the plate junction of the Nazca, Caribbean, and South American plates along the western margin of the country, where tough terrains with steep slopes dominate the landscape. These tectonic conditions result in a medium to high seismic hazard for much of the territory (Vega and Hidalgo 2016). The hillslopes are overlain by thick profiles that consist of residual soils, saprolites, and weathered rock horizons caused by deep tropical weathering (Ojeda and Donnelly 2006). Geomorphologically, the Andes Mountains that cross Colombia from south to north make the country extremely vulnerable to landslides. Additionally, the Andean Mountains influence the dynamics of weather patterns and rainfall over the region. Strong topographic features induce local atmospheric circulations that enhance deep convection, in turn leading to highly intense storms in space and time, triggering flash floods and landslides (Aristizábal and García-Aristizábal 2020; Espinoza et al. 2020; Arias et al. 2021).

Climatological and rainfall context

Colombia’s climate is tropical. In the Andean region, the mean annual precipitation ranges from 1000 to 3000 mm/year, where strong topographic features produce local atmospheric circulations and convective rainstorm events (Mesa et al. 2021; Cullen et al. 2022). The tropical climate displays a distinct pattern of intermittent rainfall, both in its spatial distribution and temporal occurrence. Bimodal regime occurs predominantly in the Andean region where rainfall peaks are prominent in the season March–April-May (MAM) and September–October-November (SON), and on the other hand, in transition zones between unimodal and bimodal regimes, a mixed regime is registered. Several authors have pointed out the meridional migration of the Intertropical Convergence Zone (ITCZ) as the main driver of the annual cycle of Colombia’s hydro climatology (Urrea et al. 2019), because the double migration of ITCZ strongly controls the bimodal regime of rainy periods. Although the ITCZ position is fundamental, is not the only factor, it does not explain all the complexity, because the migration of the ITCZ and three low-level jet streams Chocó, Caribbean, and South America are part of the complex circulation that produces high precipitation (Mesa et al. 2021).

Input data and methods

Datasets for analysis

Landslides dataset

Two landslide databases were used from the Colombian Geologic Service (SGC, in Spanish) — Mass Movement Information System (SIMMA, in Spanish, https://simma.sgc.gov.co) and the Universidad Nacional de Colombia — Geohazards research group (https://geohazards.com.co). A merge process of landslide datasets was carried out. The merging process resulted in a small number of repeated landslides which were removed. According to the geographical data collected and used in this study, 4976 landslides in the Andean region of Colombia between 1981 and 2022 were analyzed (Fig. 1c). Landslide database has information about the type of mass movement according to the update of the Varnes’ classification of landslide types which can be seen in detail in Hungr et al. (2014). According to this classification, the considered types of movements are fall, topple, slides, lateral spread, flow, and creep. Given the temporal uncertainty of landslide events in local databases, this limitation was addressed by focusing on events occurring at a minimum temporal scale of a month, and this is conservative way to deal with temporal uncertainty because sometimes the reported day could be the same or closer in time to the real occurrence date. This approach stems from the fact that a significant portion of recorded landslide events carries, at a minimum, information about the month and year of occurrence. In terms of spatial accuracy, the inventory used encompasses the centroid of the municipal boundaries or second-order administrative units, even in many cases, the exact location, situated within the Andean region. This selection criterion ensured that the identified landslide events indeed took place within the geographic scope of our study.

Rainfall dataset

The Climate Hazards Group Infrared Precipitation with Stations (CHIRPS) v2.0 dataset was chosen for conducting this study given that Institute of Hydrology, Meteorology and Environmental Studies (IDEAM, in Spanish) data was used for validating its accuracy in complex terrains of tropical regions, and this database has been used and validated in previous studies to assess the characteristics and implications of rainfall regimes in Colombia, exhibiting good performance and preserves important rainfall characteristics such as mean and seasonality in different time scales for the entire territory (Urrea et al. 2016, 2019; Fernandes et al. 2020; López-Bermeo et al. 2022; Ocampo-Marulanda et al. 2022). Comparing with rain gauge records from the IDEAM, correlations of 97% have been obtained, with mean average error of 38 mm (Mesa et al. 2021), accompanied by the benefit of relatively low latencies (Funk et al. 2015). Given that previous works concluded that CHIRPS performs well on a monthly time and spatial scales, for this study, CHIRPS monthly data was retrieved from UCSB data portal (https://www.chc.ucsb.edu/data/chirps) for the time period 1981–2022.

ENSO Index dataset

The Multivariate ENSO Index (MEI) is an index that incorporates several oceanic and atmospheric variables, including sea surface temperature (SST), sea level pressure, and zonal and meridional wind stress, to provide a more comprehensive measure of ENSO conditions. The MEI has shown to be more reliable and robust than other ENSO indices for predicting rainfall variability in many regions, including South America (Salas et al. 2020). MEI Index has been used in numerous studies to investigate the relationship between ENSO and rainfall variability (Nicholls 1988; Xu et al. 2004; Yun et al. 2021) and provides a more vivid representation of characterizing the ENSO phenomenon relative to other indices because it reflects more appropriately the nature of the coupled system between the ocean and the atmosphere integrating more information, including worldwide corrections with surface temperatures and rainfall (Mahmood et al. 2006; Baddoo et al. 2015). Thus, it is also less vulnerable to occasional technical errors in the monthly updating cycles (Naranjo Bedoya et al. 2019). Therefore, this index and the methodology for SST index of the Japan Meteorological Agency (JMA) were used to establish the prevalent ENSO phase for each year from monthly records due to the good selection of ENSO events. The MEI index v2.0 values obtained from the Physical Sciences Laboratory of the National Oceanic and Atmospheric Administration (NOAA) (https://psl.noaa.gov/enso/mei/#data) were used to quantify the ENSO variability (Wolter and Timlin 2011). Negative values of the MEI represent the cold phase of ENSO (La Niña), while positive values represent the warm phase (El Niño). Regarding the strength (intensity) of an event, and complementing what is described in Baddoo et al. (2015), absolute anomaly values until 0.4 are neutral, weak (0.5 to 0.9), moderate (1.0 to 1.4), strong (1.5 to 1.9), and very strong (≥ 2.0), and an event is said to have occurred if the threshold is equaled or exceeded for at least three consecutive overlapping 3-month periods.

Methods

The data regarding landslides and rainfall were analyzed on a monthly basis. To effectively visualize patterns over interannual timescales, both datasets were standardized and scaled using the monthly mean and standard deviation. This process helped to remove any stationary cycles from the data, resulting in a clearer representation of the overall trends. MEI ENSO Index v2.0 values were assigned to the second month of the bi-monthly time series records according to the computation details in Wolter and Timlin (1993, 1998) to consider ENSO’s seasonality. Additionally, in this study, linear and non-linear analysis methods were used, including Wavelet Spectrum Analysis (WSA), and space-time statistical methods. Relationships between landslide events, rainfall, and ENSO were estimated by aggregating the information at regional, municipal, and cluster scales. Likewise, the relationships between the studied variables were also performed by splitting the data into different temporal scales, annual, monthly, and seasonal.

Descriptive analysis and interaction with conditioning factors

While it is true that rainfall plays a significant role in determining the spatial distribution of landslides, it is important to note that other factors such as slope, geology, landcover, and soil moisture also play a critical role in their occurrence. As a result, an analysis of the local terrain settings in order to get a better understanding of the interaction between these factors and their impact on landslide occurrence was conducted. By considering these additional factors, a more comprehensive understanding of the mechanisms behind landslide occurrences can be obtained. In this study, slope gradients were derived from a 30-m spatial resolution Digital Elevation Model (DEM) from Shuttle Radar Topography Mission obtained from IGAC. Regarding geological data, re-classification proposed by Gomez et al. (2023) from the map of the lithological units from the Colombian Geological Survey (1:1,000,000) was used. The landcover types were derived from the 1:100,000 landcover map provided by IDEAM, according to the classification CORINE Land Cover Colombia. Additionally, a Mass Movements Hazard Map provided by SGC was used for comparison of landslide records in the database and landslide-prone areas in the Andean region. The soil moisture data was obtained from FLDAS Noah Land Surface Model L4 Global Monthly (MERRA-2 and CHIRPS) provided by the NASA Earth Observation data portal. The results of this analysis are shown in the supplementary material (Fig. SF1).

Wavelet spectral analysis

Time series of landslides, mean rainfall, and MEI ENSO Index v2.0 were analyzed using Wavelet Spectral Analysis (WSA). This technique allows to discover underlying periodicities in between series. WSA allow identifying time intervals and period bands in which two time series are related (Torrence and Compo 1998). Because the hydrologic series are usually complex and show non-stationary and multitemporal scale characteristics at different scales, to deal with this problem, the wavelet transformation has been widely applied in several fields of hydrology which were reviewed by Sang (2013) and related in the studies conducted by Sang et al. (2011, 2013), Rathinasamy et al. (2014), Rhif et al. (2019), and Bittner et al. (2021). Spectro-temporal methods allow for the characterization of frequency variations (Rhif et al. 2019). Continuous wavelet transform can be used to analyze the frequency and phase variations across time in a signal at several scales simultaneously, considering non-stationary power at many different frequencies (Torrence and Compo 1998; Díaz and Villegas 2022). Wavelets allow studying localized periodic behavior looking for regions of high power in the frequency-time plot. The continuous wavelet transform Tx is defined as the convolution of the time series xt with a set of “daughter” wavelets Ψ (t–τ/s) which are generated by the “mother” wavelet Ψ(t) by translation in time by τ and scaling by s:

$$T_{x} (\tau,\;s)\;=\;\sum\nolimits _{t} x_{t} \frac{1}{\sqrt{s}} \Psi^{*}(\frac{t\;-\;\tau}{s})$$
(1)

The symbol * represents the complex conjugate. The “mother” wave Ψ implemented in this study is a Morlet wavelet:

$$\Psi \left(t\right)\;=\;{\pi }^{-1/4}{e}^{i\omega t}{e}^{-{t}^{2}/2}$$
(2)

Previous research about the time-frequency evolutions of hydroclimatic series have shown that Morlet is better than others (Díaz and Villegas 2022), and it provides a good balance and performance between time and frequency domains (Le 2017). The square of the amplitude of each periodic component found in xt and how it evolves with time gives information about time-frequency wavelet energy density Px(τ,s) or wavelet power spectrum:

$${P}_{x}\;\left(\tau ,s\right)\;=\;\frac{1}{s}{\;\left|{T}_{x}(\tau,s)\right|}^{2}$$
(3)

The cross-wavelet analysis is based in the transform of the two signals in the frequency and time domain, in this case, the time series of mean rainfall, MEI ENSO Index v2.0, and landslides:

$${T}_{x,y}\;\left(\tau,s\right)\;=\;\frac{1}{s}\;{T}_{x}\;(\tau,s)\cdot {T}_{y}^{*}(\tau,s)$$
(4)

where Tx (τ, s) and Ty (τ, s) are calculated according to Eq. (1). The module of Eq. (4) is the cross-wavelet energy density and produces the cross-wavelet power spectrum. Finally, the cross-wavelet is defined as:

$$C\;\left({x}_{t},\;{y}_{t}\right)\;=\;\frac{s{T}_{x,y}}{\sqrt{s{P}_{x}s{P}_{y}}}$$
(5)

The letter s that precedes each amount indicates that these values should be smoothed. In this study, the continuous wavelet transform based on Torrence and Compo (1998) and the cross-wavelet analysis based on Maraun and Kurths (2004) were carried out using the Waipy toolkit 0.0.1.4.0 developed and implemented by Mabel Calim Costa (https://github.com/mabelcalim/waipy).

Statistical and spatiotemporal analyses

Retrospective space-time analysis with a permutation probability model was conducted to identify clusters with the open source SaTScan™ v10.1 (https://www.satscan.org) (Kulldorff 1997; Kulldorff et al. 2005). Scan statistics has been used to detect spatiotemporal clusters in other studies associated with landslides (Tonini and Cama 2019; Tonini et al. 2014, 2022). Scan statistics encompass a range of techniques that enable the identification and assessment of the statistical significance of spatial and/or space-time clusters of discrete events. In purely spatial analysis, circular moving windows are employed to scan the study area, gradually increasing their radius until reaching a predetermined fixed limit (Rmax). These windows systematically traverse the entire region of interest, counting the number of events that fall within (c) and outside (C–c) their respective areas. At each location and window size, the likelihood ratio (LR) is computed to determine the probability of a single window containing a greater number of observations than expected under a random distribution (E[c]). This calculation involves comparing the observed event count to the background population, enabling an assessment of statistical significance. The window with the highest LR is considered the most likely cluster. Under Poisson assumption, LR is computed as follows:

$$LR\;=\;{\;\left(\frac{c}{E[c]}\right)}^{c}{\;\left(\frac{C\;-\;c}{C\;-\;E[c]}\right)}^{C\;-\;c}$$
(6)

where (C-E[c]) is the expected number of cases outside the window. To assess the statistical significance of the most probable cluster, Monte Carlo hypothesis testing is employed by generating numerous random replications of the original dataset. The maximum likelihood ratio (LR) obtained is ranked to determine the significance level (p-value). This process involves comparing the observed LR to the distribution of LR values from the replicated datasets. By establishing a predetermined threshold (p-value), typically controlled by the number of Monte Carlo replications, the null hypothesis of randomness can be accepted or rejected.

When conducting space-time permutation scan statistics analysis, cylindrical scanning is employed. These scanning cylinders incorporate both the temporal dimension (t) and the geographical spatial dimension. The height of the cylinder represents the temporal aspect, while the area (s) of the circular base reflects the spatial dimension over a given radius (R) centered around each observed event. By employing these scanning cylinders, the analysis effectively captures and evaluates the space-time characteristics of the data, facilitating the identification of significant clusters. This analysis assumes that the spatial and temporal locations of all cases are independent of each other. The observed cases can be aggregated into time intervals, allowing to adjust for temporal trends and to reduce the computing timing. The expected number of events (μst) is estimated by comparing the number of observed events within a certain space-time cylinder (cst) with the expected number of events in the absence of space-time interaction. This is obtained by supposing that each event in a given area (s) observed at a time (t) has the same probability for all the times according to:

$$\mu_{st}\;=\;\frac{1}{C}(\sum\nolimits_{s}c_{st})(\sum\nolimits_{t}c_{st})$$
(7)

Then, the generalized likelihood ratio (GLR) for every space-time cylinder is computed as follows:

$$GLR\;=\;{\left(\frac{{c}_{st}}{{\mu }_{st}}\right)}^{{c}_{st}}{\;\left(\frac{C\;-\;{c}_{st}}{C\;-\;{\mu }_{st}}\right)}^{C\;-\;{c}_{st}}$$
(8)

For each potential scanning cylinder, the ratio is calculated and maximized to determine the likelihood of a cluster. The scanning cylinder that yields the highest value of the likelihood ratio (GLR) is identified as the most probable cluster, indicating that it is least likely to have occurred by chance. Subsequently, additional clusters are identified as secondary based on their respective GLR values. To assess the statistical significance of the detected clusters, a significance test is performed using Monte Carlo simulations. Afterward, the clusters are ranked according to their GLR values, with the cluster exhibiting the highest likelihood ratio considered the most probable, followed by secondary clusters with lower values. To avoid the aggregation of observed events into very large clusters, in the present study, the following values were applied, according to adopted and suggested values of other studies (Fuchs et al. 2012; Gökkaya 2016; Tonini et al. 2022): Rmax = 50 km, Tmax = 6 months, time precision = month. The inference was performed by applying a Monte Carlo simulation with 999 permutations. The significance level was defined as α = 0.05 (significant).

Finally, spatial patterns in aggregated data at the municipal level were analyzed by spatial statistics, using landslide densities and hot spot (cold spot) analysis test using the Getis-Ord Gi* statistical to identify the location of local clusters for landslide occurrence. In the Getis-Ord Gi* method (Getis and Ord 1992), a local weighted mean around each observation including the observation itself is compared to the mean of the entire dataset (Gökkaya 2016). The analysis specifies the spatial relationship of the feature to neighboring features and a threshold distance beyond which adjacent features are ignored. A fixed distance band spatial relationship with Euclidean distance and a 50-km threshold distance were used in this analysis. The landslide densities and hot spot (cold spot) analysis were performed in ArcGIS 10.7 software (https://www.esri.com). The results of this municipal scale analysis are shown in the supplementary material (Fig. SF9).

Results

Landslides and rainfall variability

In this study, 4976 landslides were analyzed between the years 1981 and 2022 in the Andean region of Colombia, where 0.7% corresponded to creeps, 6.2% to falls, 6.1% to flows, 86.3% to slides, 0.5% to spreads, and 0.2% to topples. Many of these movements resulted in several fatalities and economic losses. In the region, a marked bimodal rainfall regime cycle was observed, with maxima in total events in the seasons MAM (1540) and SON (1296), and minima in DJF (924) and JJA (1216). Figure 2 shows landslide types and mean total annual rainfall. Landslides were mainly oriented following the extension of the mountain ranges, with the highest occurrences in the years 2010–2012, 2017, 2020–2022, mainly in departments such as Antioquia (13.7%), Caldas (12.8%), and Cauca (12.6%), corresponding with some of the departments with more economic affectations according to different databases and studies (DesInventar, EM-DAT, GLC-NASA).

Fig. 2
figure 2

Spatial distribution of landslide types and bar chart of total landslides with the mean total annual rainfall in the Colombian Andean region in the period 1981–2022

Full size image

Other secondary concentrations of landslide occurrence were observed in departments such as Boyacá, Norte de Santander, Tolima, and Nariño, which together account for about 27.0% of the analyzed records. Similar trends were observed in rainfall at seasonal scale (Fig. 3), with a predominant increase in occurrences at the MAM and SON seasons, corresponding with the rainfall peaks in the region. Monthly rainfall was slightly higher for the rainy season SON than MAM, and lower in the dry season DJF. However, the season JJA showed higher values of rainfall for this dry season. According to the type of landslide, accounting for a 52.4% of the total records, major slides were observed in the departments of Antioquia (15.3%), Caldas (15.2%), Cauca (14.5%), and Boyacá (7.4%). In terms of flows, a marked pattern was found in the departments of Tolima (28.5%), Caldas (27.2%), Boyacá (13.6%), and Norte de Santander (8.9%). Regarding falls, the departments of Caldas (14.6%), Boyacá (14.2%), Cauca (11.0%), and Norte de Santander (9.4%) presented the highest number of occurrences, with an approximately 49.2% of the total records. About creeps, spreads, and topples, the number of records in the databases was very few compared to the other types, presenting high spatial dispersion without clearly noticing patterns. In relation to the occurrence of landslides according to the type and season, it can be deduced that in terms of slides, they occurred predominantly in MAM (32.9%) and SON (26.7%), while flows predominated in DJF (43.4%) and falls in JJA (42.1%). The remain movements presented a more uniform seasonal distribution.

Fig. 3
figure 3

Spatial distribution of landslide types and bar chart of seasonal landslides with the mean total seasonal rainfall in the Colombian Andean region in the period 1981–2022

Full size image

To analyze the relationship between the annual cycle of rainfall and landslides, the mean monthly rainfall with the number of landslides that occurred in each month was compared (Figs. 4 and 5). The highest number of landslides usually occurred in April–May and October–November, which were the months with the highest rainfall intensity in many areas of the region. In all months, the physiographic influence of the Colombian territory can be observed, mainly dominated by the three branches of the Andes Mountain range. On a monthly temporal scale, it can be seen that landslide records were mostly coincident with the highest mean monthly rainfall. This was more noticeable from the year 2005, the period 2010–2011, 2017, and 2020–2022. Nevertheless, considering the different types of analyzed events, in terms of creeps, spreads, and topples, they presented a fairly uniform behavior throughout the year, without showing a predominant trend. Falls presented a unimodal behavior with a maximum peak in the month of July. As for flows and slides, their behavior was bimodal, with maximum peaks in the months of July–December and May–November, respectively (Fig. SF2).

Fig. 4
figure 4

Spatial distribution of landslides and rainfall in the Colombian Andean region: Mean total monthly rainfall and number of landslides events in the period 1981–2022

Full size image
Fig. 5
figure 5

Bar charts of monthly landslides and the mean monthly rainfall in the Colombian Andean region (1981–2022)

Full size image

ENSO’s effects on landslides

To consider the effect of the ENSO on the occurrence of landslides, the years of the time series were classified in Neutral, El Niño, and La Niña years. Figure 6a shows the standardized MEI ENSO index v2.0, standardized mean annual rainfall, and landslides time series in the region. It was noticeable that the peaks of landslide coincided with negative values of the MEI index, corresponding mainly with the strong periods of the La Niña, where the wet seasons in Colombia become not only more intense but also longer (Poveda et al. 2014). This was more remarkable from the events recorded from the year 2000 onwards. The highest annual landslide occurrence periods corresponded to 2010–2011 and 2020–2022 when La Niña years were reported in the analyzed databases, corresponding with the rainiest years 2011 (2681 mm) and 2022 (2734 mm) in average over the study region. It is worth noting some landslide peaks in La Niña periods were preceded by strong El Niño periods, such as 1987–1988, 1997–1998, and 2015–2016. Figure 6b to d shows the spatial distribution of rainfall and landslides for the three categories. The highest number of landslide recurrences occurred in La Niña years (56.7%), while the Neutral years account for 36.5%, and finally the El Niño years with 6.8%. It should be noticed that during the study period, only 12 years were classified as La Niña, 10 as El Niño, and 20 as Neutral. These findings underscored the substantial impact of climate variability on landslides, emphasizing the non-linearity inherent in the relationships among landslides, rainfall patterns, and ENSO.

Fig. 6
figure 6

a Temporal distribution of landslides in the Colombian Andean region between 1981 and 2022, Multivariate ENSO Index v2.0 and mean annual rainfall. Spatial distribution of landslides considering ENSO years: b Neutral, c El Niño, and d La Niña

Full size image

Figure 7a shows the spatial distribution of landslides according to ENSO phase intensity. In Fig. 7b can be noticed that more than 63% of landslides were related to La Niña phase years, distributed mainly along the western and eastern mountain ranges. Figure 7c shows the average annual cycle for landslides and rainfall for the three categories. Again, it was observed that the landslide occurrence in the La Niña phase dominated, with monthly percentages in general higher than 50%, and this difference was more noticeable in dry periods. The monthly variation of rainfall and landslides for each category showed a bimodal behavior, with peaks coinciding in the months of April–May and October–November. The bimodal behavior between rainfall and landslide records was more similar and accentuated for Neutral years and less during La Niña and El Niño years.

Fig. 7
figure 7

a Spatial distribution of landslides according to ENSO phase intensity. b Percentual distribution of landslides according to ENSO phase (LN, La Niña; EN, El Niño) intensity (VS, Very Strong; S, Strong; M, Moderate; W, Weak). c Average monthly rainfall cycle (dashed lines) and registered landslides (bars) in the Colombian Andean region for each phase: El Niño (red), La Niña (blue), Neutral (green)

Full size image

Figures 8 and 9 show the distribution of landslides and rainfall considering ENSO La Niña years, the most critical phase in terms of landslide occurrence in the region. For a spatial detail in El Niño and Neutral years, please refer to supplementary material (Figs. SF3 to SF6). Additionally, it was analyzed the landslide type regarding the MEI Index v2.0, where can be noticed the effect of La Niña phase, which had an incidence in more than 80% of topples and falls, around 60% in slides, around 55% in flows and creeps, and a remarkable 95% of spreads. Regarding intensity, very strong La Niña had the highest effect on creeps, moderate La Niña had the highest effect on topples, and moderate and weak La Niña on spreads and falls, while remain similar for other types. Slides, flows, and creeps had quite similar effect of Neutral phase and El Niño ENSO intensities (Fig. SF7).

Fig. 8
figure 8

Spatial distribution of landslides and rainfall in the Colombian Andean region considering ENSO La Niña years: Mean monthly rainfall and landslides in the period 1981–2022

Full size image
Fig. 9
figure 9

Bar charts of monthly landslides and the mean monthly rainfall considering ENSO La Niña years in the Colombian Andean region (1981–2022)

Full size image

The annual rainfall cycle of El Niño years retained bimodality but is characterized by wide valleys (February to April and July to September) and sharp peaks in the months of June and October, which means more marked transitions between rainy periods and dry periods. The annual rainfall cycle of La Niña years was characterized by broad peaks, with valleys in the months of January–February and August–September, with smooth transitions that reflected a smoother bimodal cycle. The results indicated that the difference between the annual rainfall cycle according to the ENSO was not only in the rainfall peaks, but also in the presence of more rainfall in the dry season months at the middle and end of each year (Naranjo Bedoya et al. 2019), although some consistent anomalous features in the dry season are not discussed since they do not mean really heavy precipitation (Grimm and Tedeschi 2009).

Landslides responded to changes generated by the influence of ENSO. In general, for the rainy periods in La Niña, the number of landslides was higher than the multiannual average in the study period. Similarly, in dry periods, the number of landslides was higher than the multiannual average. Figures 6 and 7 provide valuable insights into the effect of ENSO on landslide occurrence, exhibiting variations across different ENSO years. Notably, during El Niño, landslides reached their minimum value in dry periods, with the lowest occurrence observed in August. In Neutral years, the annual rainfall cycle behaved similarly to La Niña during peaks in the first rainy season and to El Niño during the transition to the second rainy season. Regarding landslide records, the October peak disappeared, replaced by higher peaks in May and November, with slightly lower valleys observed in August–September and January–February.

Analyzing the data from Figs. 6 and 7, it becomes evident that the effects of ENSO on landslide occurrence differ between ENSO years. For the rainy seasons (MAM and SON), rainfall was higher during La Niña years compared to El Niño years. This difference became even more pronounced during the dry seasons (DJF and JJA), where rainfall in La Niña years was considerably higher. In JJA, the dry season, the mean monthly rainfall during La Niña years ranged between 150 and 200 mm, while in DJF, the dry season, the mean monthly rainfall was around 100 mm (Figs. 8 and 9).

Relationship between rainfall, ENSO, and landslides

Frequency domain analysis

The linkages between atmospheric modes and the periodicity in landslides and rainfall were explored using wavelet and Fourier analysis. Spectral analyses were conducted for monthly data observed over a 42-year period records. Before WSA, the landslide time series was normalized to monthly landslide rates by the annual totals to reduce gaps into the spectrum, improving the interpretability of the results. Figures 10a–c show to the right side, the wavelet analysis of the time series, decomposing them into time and frequency space, while on the left side, the global wavelet spectrum (including wavelets and Fourier spectrums and 95% confidence line of normalized landslide events, mean monthly rainfall in the Colombian Andean region, and MEI ENSO Index v2.0 is presented).

Fig. 10
figure 10

Wavelet power spectrums and time-integrated global power spectrums (blue lines), Fourier spectrum (gray lines), and the 95% confidence interval (red lines) for: a landslide time series in the Andean Mountain region. b Monthly rainfall. c Multivariate ENSO Index v2.0. d Cross-power wavelet spectrums of rainfall and landslides. e Cross-power wavelet spectrums of Multivariate ENSO Index v2.0 and landslides. The arrows indicate the relative phase relationship between the two variables, with in-phase pointing right. Non-hatched zone indicates the significant cone of influence

Full size image

In Fig. 10a, the response of landslides to the modes of rainfall variability, annual cycle (Fig. 10b), and ENSO (Fig. 10c) can be noticed. The bimodal behavior of rainfall can be noticed through the main peaks at 0.5- and 1-year period in the wavelet spectrum. This situation reflected the annual cycle of rainfall that predominated. The power on these scales stood out more than the others, explaining most of the rainfall variance by 0.5-year periodicities, followed by a period of about 1 year, with high power for 1985, 2010–2012, and 2017 onward, corresponding with the peaks detected in Fig. 2a. These time intervals coincided with some of the largest number of landslide events in the region mainly occurring during Strong La Niña 2010–2011 and 2020–2022. Other periodicities with high power can be found between periods of 3 and 4 years, mainly between 1998–2000 and 2010–2015. These periods evidenced the strong relation between landslides events and rainfall. Other areas within low to moderate frequency bands were also visible in the spectrum of Fig. 10a, which could correspond to ENSO with other quasi-decadal oscillatory processes (Díaz and Villegas 2022). Figure 10b shows higher power zones around the 4- and 6-year period between 1998 and 2000, confirming the presence of strong La Niña events in these years.

Figures 10d and e show the cross-wavelet transform between the analyzed time series. On the upper side, the amplitude of the standardized time series signals is presented, while on the lower side, the cross-power wavelet spectrums of the time series, with arrows indicating the relative phase relationship between the variables, are shown. In Fig. 10d, the greatest power between rainfall and landslide time series was observed following the variability mode of the annual rainfall cycle, especially for the first half of years (0.5-year period), where both variables were pointing almost in-phase. Also, for a 3- and 4-year period, both variables showed certain in-phase behavior from 1998–2000 to 2010–2014. The aforementioned periods can be associated with the quasi-periodic ENSO occurrence and its extreme phases, which constitutes the most important modulator of interannual climate variability affecting the annual cycle of hydroclimatic variability in tropical South America (Salas et al. 2020). In Fig. 10e, in general terms, areas of high power can be seen within the cone of statistical confidence. Nevertheless, for periods of 3 and 4 years, it can be noted that approximately between 1996–1997, 2002, 2009, and 2015, low power was observed and there was a tendency for both series to be in almost contra-phase. This was due to the occurrence of the El Niño phase in the aforementioned years. In terms of amplitude in the landslide time series, in 2010–2011, 2017, and 2020–2022, the major peaks were presented, coinciding mainly with moderate and strong episodes of La Niña phase.

Space-time analysis

From a space-time perspective, in this study, the spatial pattern of landslides recorded in the Colombian Andean region was analyzed. It resulted to be globally clustered both in space and in time. For the analyses, landslide events were treated as punctual data, under the assumption of spatial inhomogeneity of the patterns, as naturally is for geological events (Tonini et al. 2014). Several clusters in different zones of the region emerged, which were indicated as circles of varying sizes in Fig. 11. The results of the analysis reflected the interaction between geographical position along the mountain ranges and rainfall variability. The mean total multiannual rainfall in the areas grouped by clusters was around 2100 mm (refers to Fig. SF8 for details of detected clusters and mean total annual rainfall). Table 1 shows the results of detected clusters for the 42 years of analyzed landslides, of which 42 landslides clusters were detected, but only 34 of them showed statistical significance (p-value < 0.05). Sixty-five percent of the significant clusters occurred in La Niña years, followed by 26% in Neutral years and the remaining 9% in El Niño years (Fig. 11a and b).

Fig. 11
figure 11

a Total number of detected clusters with respective p-values. b–i Statistically significant space-time clusters of landslides in the Colombian Andean region, over the average annual rainfall of the analyzed period. The size of the circles indicates the spatial coverage of the detected landslide clusters. j Detected cluster with its number of observed landslide events and cluster radius. The bar chart color indicates the ENSO phase

Full size image
Table 1 Characteristics of the clusters detected by the multiannual analyses between 1981 and 2022 in the Colombian Andean region, via the scan statistics permutation model
Full size table

Figures 11c–i show that the clusters detected initially occurred in the western mountain range until 2004 (departments of Antioquia, Caldas, Nariño, and Risaralda). Later, clusters were detected in the eastern mountain range (department of Norte de Santander) and from 2010 onwards, in the central mountain range (departments of Huila, Cauca, and Tolima). For the periods 2011–2016 and 2017–2022, the presence of clusters was evidenced in mountainous territories of the three mountain ranges, including occurrences in the departments of Cundinamarca, Santander, Boyacá, Valle del Cauca, and Quindío, in addition to the usual occurrences already mentioned and detected as clusters in previous analysis periods. Also was noticed a notable occurrence of landslide clusters corresponding with La Niña and Neutral years for the considered 6-year periods. Bar charts of Fig. 11j count the observed clustered landslides according to the ENSO phase.

Discussion

The connection between rainfall and landslides in Colombia has been extensively studied and established, especially in the Andean region. However, there is room to delve deeper into space-time analyses across various scales and account for the impact of ENSO. This has been explored in some studies as Naranjo Bedoya et al. (2019), Emberson et al. (2021), Garcia-Delgado et al. (2022), and Gomez et al. (2023). Notably, as previous studies, this study highlights the Andean region as an area with high landslide risk in Colombia. This risk is further exacerbated by projected changes in rainfall, as reported by IDEAM, which suggest a range of 10 to 40% increase in rainfall at a 2040 horizon in most parts of this region associated to climate variability. The likelihood of exceeding the stated rainfall thresholds is especially probable in this area, where 70% of the territory is exposed to 2000 mm of rainfall per year (Gómez et al. 2021) and landslide probabilities exceed 60% (Gómez et al. 2023a, b), all of this, without considering other effects related to ENSO. The findings of this study underscore the urgency of understanding landslide risk in the region, considering the potential impacts of changing rainfall patterns in the near future.

In this study was noticed, at Andean region scale, a bimodal landslide distribution with a bimodal annual rainfall cycle. A noteworthy observation of this study was that in 98.1% of the analyzed cases (4880 events), the mean annual rainfall in the Andean region surpassed the critical rainfall threshold calculated using the proposal of González-García and Mayorga-Márquez (2004). Furthermore, 18.6% of the 4976 compiled events exceeded this critical rainfall threshold and were associated with rainfall as triggering factor in the database. These findings emphasize the importance of rainfall as a critical factor in landslides in the region, and further research may help uncover the underlying mechanisms and implications of these relationships. The rainfall monthly cycle in the study area presented a bimodal behavior characteristic found in other studies (Urrea et al. 2019; Mesa et al. 2021), with maximum percentage values of rainfall in the seasons of MAM and SON (Fig. 3). The MAM rainy season was temporally shorter with a higher value of accumulated rainfalls, which means less rainy days but more rainfall volume, so higher intensities. The second rainfall season, SON, showed more rainy days with less rainfall volume (Aristizábal and García-Aristizábal 2020).

ENSO affects the amounts and mode of distribution of rainfall (Poveda et al. 2011, 2002). This study confirmed the strong direct influence of rainfall and ENSO on the occurrence of mass movements, which coincided with the results reported by other studies in some areas of the region (Klimeš and Escobar 2010; Sepúlveda and Petley 2015; Naranjo Bedoya et al. 2019; Gómez et al. 2021; Garcia-Delgado et al. 2022; Gomez et al. 2023). A slightly different number of landslides registered in the second half of the year, July to December (2489 events) related to the first half of the year, January to June (2487 events) was noticed; this was correlated directly with the second rainy season of the year (SON). A noteworthy observation pertained to the lag correlation between the peaks in landslides occurring during La Niña ENSO phase periods and the corresponding peaks in mean rainfall. This correlation was particularly noticeable for the second peak in November (Fig. 7c). Garcia-Delgado et al. (2022) attributed this to the long duration of the effect on rainfall totals during La Niña phase events, possibly enhancing longer soil saturation for several weeks after the rainfall peak. Nevertheless, during ENSO-Neutral years, the time lag in landslide occurrences similarly manifested during the second rainfall peak. In contrast, during El Niño periods, the time lag was observed in conjunction with the first peak. This could also be seen in Klimeš and Escobar (2010) and Naranjo Bedoya et al. (2019) results for the Northeast and Western region of the department of Antioquia and the La Iguaná Valley, respectively, in the month of June. The variability in these time lags highlights the complex and non-linear nature of climate-landslide relationships during different ENSO phases and suggest that in ENSO-Neutral years, the time lag in landslide occurrence aligning with the second rainfall peak may be influenced by the persistence or cumulative effect of rainfall over time. This lag could be related to the gradual saturation of soils, increased groundwater levels, or other hydrological processes that contribute to landslide susceptibility. The relatively stable climate conditions during ENSO-Neutral periods may allow for a delayed response to the second peak in rainfall. In El Niño phase periods, the time lag occurring with the first rainfall peak may be a result of the rapid onset of heavy rainfall. The initial peak in rainfall could saturate soils quickly, leading to a higher likelihood of landslides in the immediate aftermath of intense rainfall events. The heightened sensitivity to the first peak may be attributed to the more dynamic and extreme weather conditions associated with El Niño. Finally, the results suggest that in La Niña phase periods, the time lag in landslides might exhibit its own distinctive pattern because periods of recurring rainfall events, contributing to soil saturation over time.

The second half of the year exhibited a 1-month displacement of landslide peak in November with respect to the rainfall peak of October (Figs. 4 and SF2). The increase in landslides in November indicated a relationship between landslide and antecedent rainfall, associated with the increase in soil moisture. This trend could reflect the influence of antecedent rainfall on the occurrence of landslides after a dry season in the middle of the year. This condition was noticeable independently of the ENSO year type. Previous water accumulation increases the water content within the soils, increasing the landslide susceptibility during subsequent low monthly intensities (Gómez et al. 2023b). The antecedent soil wetness condition plays a crucial role in the occurrence of rainfall-induced landslides. This is particularly important considering that both long-term accumulated rainfall and the recent rainfall condition contribute to the occurrence of landslides (Zhao et al. 2019). However, antecedent rainfall of up to 30 days is more relevant than daily intensity in its occurrence (Gomez et al. 2023). Some prominent cases related to this can be related to the events occurred on the Medellín-Bogotá Highway in 2016, Dabeiba-Antioquia in 2020, and the Panamerican Highway in Rosas-Cauca in 2023. This can be attributed to the fact that tropical soils with deep weathering profiles, as is the case of the Colombian Andean region, are associated with longer periods of rainfall and not with particular events of high intensities (Gómez et al. 2023a). Analogously, high rate of landslides in June and July, a dry season, indicated a relationship between landslide and antecedent rainfall, associated with the increase in soil moisture from the first rainy season. This can explain the no direct correlation between rainfall and landslides in this season, noticed for the department of Antioquia in the results of Aristizábal and García-Aristizábal (2020).

Analyzing the used database can be noticed that few landslides were registered in the 80s and 90s (Fig. 2). In this period, Garcia-Delgado et al. (2022) highlighted that the trend in fatality-inducing landslides had a decrease in the median values in relation to previous years, despite the fact that two of the most catastrophic events occurred in this period, Armero-Tolima in 1985 and Páez-Cauca in 1994. They also suggested that part of the explanation is associated with the implementation of the National System for Disaster Attention and Prevention, according to Law 46 in 1988, but later questioned its effectiveness given the increase in the median number of fatal events in the two subsequent decades. On the other hand, most of the analyzed landslides in the used database occurred after 2000. Some landslide peaks started to appear after 2005 and considerably increased in 2010. This was coherent with the findings of Garcia-Delgado et al. (2022) who highlighted the period 2001–2010 with the highest fatal events. The increased tendency from 2000 onward could be associated to several reasons, greater tendency for government entities to register those landslides, which did not exist previously (Aristizábal and García-Aristizábal 2020), the abnormally strong winter season in Colombia in this period time, and a change in local population from a predominantly rural to an urban territory occupation, exacerbated by the issue of violence in recent decades. In fact, forced displacement in Colombia has generated flows of migrants to metropolitan areas and urban centers, thus influencing the reactivation of the rural-urban flow of internal mobility (Ruiz 2011), involving an urban expansion along landslide-prone areas in the region. The aforementioned spatial redistribution of the rural population due to demographic growth induced rapid deforestation (Lambin and Geist 2003), which has been identified as a trigger for shallow landslide activity (Glade 2003). This makes sense, considering that according to Garcia-Delgado et al. (2022), the IDEAM in 2017 indicated that the Colombian Andean region ranks as the second region in terms of deforestation in the country behind the Amazon. These reasons are in line with what was concluded years ago by Ojeda and Donnelly (2006), regarding the apparent increase in landslides in the past few decades, attributed among other reasons, to increasing migration of the population to the towns and cities.

The ENSO events have been shown in previous studies to be related to regional extremes in weather (Hanley et al. 2003). However, the impact of ENSO on extreme rainfall is weaker than on total rainfall in critical areas in Colombia, where Emberson et al. (2021) modeled significant relationships between ENSO and landslide exposure. With the aim to evaluate the potential correlation between rainfall variability and landslides with ENSO phases, in this study, an ENSO index that quantifies its magnitude was used. For instance, strong winter seasons, mainly during the La Niña years, are crucial for understanding landslide occurrences, as confirmed in Figs. 6 and 7. La Niña phase conditions may lead to increased runoff and erosion, influencing the timing of landslide occurrences during specific rainfall peaks. Moving on to Fig. 7, rainfall displayed a clear bimodal pattern independent of the ENSO phase. This behavior was also evident in landslide occurrences but was more pronounced in El Niño and Neutral years, while La Niña years exhibited a smoother pattern. These observations provided valuable insights into the relationship between ENSO phases, rainfall patterns, and landslide events, contributing to the understanding of landslide susceptibility in the region. The findings further highlight the importance of considering the specific ENSO phase when studying landslide occurrences and rainfall patterns.

Additionally, Fig. 7c shows that during the first four months of the year, a greater number of landslides were observed for La Niña years and fewer landslides associated with El Niño years. Subsequently, there was an increase in landslides for the months of May, June, and July compared to the Neutral years. From the results in Fig. 7c, the peak of landslides in the months of June and July were associated with the highest values of rainfall in La Niña years. According to Aristizábal and García-Aristizábal (2020), the notably larger rainfalls during the JJA dry season in the La Niña years compared to the El Niño years could explain the landslide concentration in the second rainy season SON because soils are in a more saturated condition before the start of the rainy season. Larger rainfall durations induce more infiltration within the soil, thus, larger water contents, which leads to reduced suction and slope stability. During the La Niña phase, landslides were more frequent and occurred throughout the year, while in El Niño years, they occurred especially in the months of May–June and October, when coinciding with rainfall peaks. However, the study conducted by Klimeš and Escobar (2010) showed a shift in the timing of peak landslide occurrence during the later rainy period during El Niño years.

The obtained results indicated differences between the phases of the ENSO in the behavior of rainfall and therefore in landslide events at the annual and monthly scales. In the El Niño phase, the bimodal behavior of the annual cycle of rainfall was more marked, with periods of strong transition. Nevertheless, in this phase, less landslides occurred according to the analysis. On the other hand, the behavior of rainfall in the La Niña rainfall phase had broadly shaped peaks, which means considerable and constant rainfall throughout the year, causing most landslides. In general terms, the second rainy season records of rainfall were higher than the first season records, and this was consequent with landslide records in analyzed database and other studies in the region previously cited, mainly during the La Niña phase, because of the coupling of the ITCZ with the dynamics of the Chocó jet and mesoscale convective systems over the Pacific coast, and the passage of tropical waves from the east during the autumn and boreal summer (Naranjo Bedoya et al. 2019).

It is reasonable to expect the frequency of extreme rainfall events to be modulated by ENSO in some preferred locations. However, this does not mean that regions with ENSO-sensitive monthly and seasonal rainfall are necessarily regions with ENSO-sensitive frequency and intensity of extreme events, and vice versa (Grimm and Tedeschi 2009). Even with identical amounts, the climate can be very different if the frequency and intensity of rainfall differ (Trenberth 2011). From Figs. 6 and 7 can be appreciated that the location of significant areas with landslides coincided with strong or very strong La Niña events. Despite this, there was a non-linear ENSO-rainfall relationship because ENSO events with similar characteristics generate different responses in rainfall and landslide occurrences. Moreover, the influence also changes depending on the site as can be seen in the differences between Figs. SF3 and SF6 (supplementary material), showing that the relationship was more complicated than simple linear correlation, with different responses observed depending on the geographical location, the characteristics of ENSO phases, and the interplay of the conditioning factors. The observed time lags during ENSO phases reflect the nuanced interactions between this phenomena, rainfall patterns, and the underlying geological and hydrological conditions that influence landslide occurrences in the region. To analyze this, it is also important to highlight how landslide peaks in La Niña periods preceded by strong El Niño periods, exemplified by noteworthy instances such as 1987–1988, 1997–1998, and 2015–2016, can be related to the occurrence of torrential flow cases as documented in Aristizábal et al. (2020), for example, events occurred in 1999 and especially for 2017, such as the impactful incident in Mocoa-Putumayo, which was discussed by Garcia-Delgado et al. (2022), regarding the incidence or not of ENSO over the effect of short-duration high-intensity rainfall at the time of occurrence. Similarly, the case for the event in Salgar-Antioquia in 2015, but for a period with negative monthly precipitation anomalies associated with El Niño conditions (Hoyos et al. 2019; Velásquez et al. 2018).

In both Fig. 10b and c can be noticed peaks in a close 10-year period remarking the aforementioned effect of ENSO’s oscillatory in rainfall patterns and the impact of interdecadal oscillations on the frequency of extreme precipitation events over South America (Grimm et al. 2016). Also, in Fig. 10d, an in-phase relationship was indicated by arrows that point straight to the right. Nonetheless, other cases showed a lead/lag relationship when rainfall according to the ENSO phase led to a rainfall response. Based in phase difference and their interpretation proposed by Torrence and Compo cited by Díaz and Villegas (2022), time series of landslides and rainfall were approximately in-phase, either with leading and lagging in time or/and frequency, from approximately 1995 to 2015 within the significant cone of influence for 3- and 5-year period which can be associated with the quasi-periodic ENSO occurrence, which modulate the interannual climate variability in Colombia. This goes in line with Baddoo et al. (2015) in relation to ENSO’s largest contribution to interannual climate variability globally with a 2–7-year period. Regarding MEI ENSO Index v2.0, Fig. 10e displays marked areas of low power within the cone of statistical confidence. A quasi-contra-phase trend between the two series for periods of 2 and 4 years can be observed, particularly for 1997, 2009, and 2015, and to a lesser extent, around the first 90 years. Notably, these time slots coincided with some of the most significant El Niño events, represented by high positive MEI Index values but with less landslide occurrences. This suggests a potential relationship between the less intensive rainfall patterns and El Niño occurrences during these periods. Further research could delve into the underlying factors behind these observed trends and explore the implications for understanding and predicting weather patterns in the region.

Prior to this study, only a limited number of studies had been concentrated on the cluster analysis of landslides in both spatial and temporal dimensions. In Colombia can be found the comprehensive study of Garcia-Delgado et al. (2022), which with another focus related with landslide fatalities and following an approach as used by Görüm and Fidan (2021) and Haque et al. (2016), found similar or comparable patterns mainly in northwestern, south and center-western of Andean region. The scarcity of research can be attributed, in part, to the absence of comprehensive or depurated databases, and therefore the analyses have been focused mainly on spatial distribution, often neglecting the temporal aspect. Despite the issue of uncertainty and incompleteness of the inventories, some studies have dealt with the issue (Guzzetti et al. 2008; Kirschbaum et al. 2015; Naranjo Bedoya et al. 2019; Gómez et al. 2021; Görüm and Fidan 2021; Garcia-Delgado et al. 2022) and even demonstrated that they can be used to determine landslide susceptibility with a reasonable degree of reliability (Klimeš and Escobar 2010). Time of landslide events is usually overlooked in territorial-scale analyses, despite space-time analysis helps identify regions with recurrent landslide activity and clusters of events beyond random distribution, possibly triggered by a single extreme rainfall event (Tonini et al. 2022).

The analysis of landslide records in this study revealed interesting patterns regarding the triggering factors. Surprisingly, only approximately 20% of the records in the database showed rainfall as the trigger for landslides. Around 60% of the events had no record of triggering factor, 17% of the events had unknown or unidentified triggers, leaving the remaining 3% to other causes but the same related to rainfall such as erosion, undermining, among others. A recurring scenario, evident in the number of comprehensive records of landslide triggers, was noted in well-documented studies by Gómez et al. (2021) and Garcia-Delgado et al. (2022). Despite utilizing different databases, both studies arrived at similar proportions of events lacking trigger documentation, 60.3% and 63%, respectively. In Fig. SF2 in supplementary material, it was compared the plotting of the landslide events with unknown trigger and the discernible significant trend indicated that the majority of those landslide events were likely induced by rainfall with the same bimodal behavior with peaks in May and November.

In this paper, initially for the analyses, landslide events were considered punctual data according to databases information, under the assumption of spatial inhomogeneity of the patterns. Nevertheless, landslides in the region were mainly triggered by hydrometeorological factors as rainfall and soil moisture, and the results clearly indicated that landslides were not randomly distributed in the study region but spatially grouped in clusters with a mean total multiannual rainfall over 2100 mm. This cluster behavior was demonstrated over a huge number of simulation envelopes through Monte Carlo test. Focusing on the landslides grouped in the detected clusters, 22% of the 3012 landslides within the 42 clusters (from 1981 to 2022) were triggered by rainfall. The average rainfall at these landslide locations was measured at 2178 mm. For the statistically significant landslides associated with the clusters, of the 2872 events, 22% were still linked to rainfall as the trigger. However, given the clarification of the trigger factor in the previous paragraph and the results shown in Fig. SF2, it was very likely that the triggering of landslides within the clusters is almost entirely associated with rainfall. Moreover, 65% of the significant clusters occurred in La Niña years, followed by 26% in Neutral years and the remaining 9% in El Niño years (Fig. 11a and j). Out of the total number of clusters identified, 11.9% were attributed to Strong La Niña events, while 33.3% were associated with Moderate La Niña events, which was consistent with previous analysis that indicated that most of landslide occurrences in rainy periods of La Niña phenomenon, mainly 2010–2011 and 2020–2022. Conversely, 7.1% of the clusters were linked to Strong El Niño events, and 4.8% were connected to Moderate El Niño events. The remaining clusters were observed during weak ENSO events and Neutral years.

Additional analysis may be required to thoroughly assess the cause-and-effect relationships between landslides and some conditioning factors. However, through careful examination of local space–time clusters, it becomes possible to identify a correlation with local meteorological conditions. Detected clusters may be attributed to the intensified occurrence of extreme rainfall events in the area during recent periods, potentially resulting from climate change, soils texture and structure, land use/cover, among others. When examining the descriptive statistics related to cluster duration on a global scale, it was observed that the number of detected clusters increases over time with variability in their duration. Nevertheless, it can be noticed that despite the most of cluster’s duration corresponded to La Niña years (as clusters # 4 and 2), two clusters (# 15 and 21) were related to a transition phase of El Niño year to Neutral year, with the highest duration of 6 months. The mean value of duration of cluster was 85 days, which means almost 3 months that can be associated with a trimonthly period. Hence, although the detected clusters occurred throughout the year, they followed a bimodal trend with peaks in MAM and SON, coinciding with the wet seasons in the region.

Conclusions

This study aimed to investigate the correlation between rainfall, ENSO, and landslides in the region. Using data from the period of 1981–2022, the relationships at different temporal and spatial scales were analyzed. According to the results, the major conclusions are:

  1. 1.

    The occurrence of landslides in Colombia is closely related to rainfall patterns, with a bimodal annual rainfall cycle observed. Peaks in landslide activity coincide with the seasons of high rainfall intensity, such as the months of MAM and SON seasons. These patterns are influenced by the El Niño-Southern Oscillation (ENSO). La Niña years, characterized by constant and considerable rainfall throughout the year, are associated with higher landslide activity. On the other hand, El Niño years exhibit a bimodal rainfall pattern, with landslide occurrences concentrated in specific months. Thus, the relationship between ENSO phases, rainfall, and landslides is complex and non-linear, showcasing varied responses depending on the location and characteristics of ENSO events. The observed time lags during ENSO phases highlight the nuanced interactions among this phenomenon, rainfall patterns, and the underlying geological and hydrological conditions influencing landslide occurrences in the Andean region.

  2. 2.

    The increase in landslides during November indicates a relationship between landslides and antecedent rainfall, which leads to an increase in soil moisture. Similarly, a high rate of landslides in June and July, during the dry season, suggests a correlation between landslides and antecedent rainfall, associated with increased soil moisture from the first rainy season.

  3. 3.

    The analysis conducted in this study has revealed the mutual complementarity of the two space-time approaches under consideration, each offering unique insights derived from landslide inventories. The implementation of kernel density maps allowed for the identification of regions exhibiting consistently high landslide density across the entire analysis period. In contrast, spatiotemporal cluster analysis identified specific areas characterized by significant and recurrent landslide activity within distinct time frames. Together, these approaches provided a comprehensive understanding of landslide patterns, highlighting both persistent high-risk regions and temporally concentrated areas of landslide occurrence particularly in changing climate times.