Satellite and hydrological model-based technologies provide estimates of rainfall and soil moisture over larger spatial scales and now cover multiple decades, sufficient to explore their value for the development of landslide early warning systems in data-scarce regions. In this study, we used statistical metrics to compare gauge-based and satellite-based precipitation products and assess their performance in landslide hazard assessment and warning in Rwanda. Similarly, the value of high-resolution satellite and hydrological model-derived soil moisture was compared to in situ soil moisture observations at Rwandan weather station sites. Based on statistical indicators, rainfall data from Integrated Multi-Satellite Retrievals for Global Precipitation Measurement (GPM_IMERG) showed the highest skill in reproducing the main spatiotemporal precipitation patterns at the study sites in Rwanda. Similarly, the satellite- and model-derived soil moisture time series broadly reproduce the most important trends of in situ soil moisture observations. We evaluated two categories of landslide meteorological triggering conditions from IMERG satellite precipitation: first, the maximum rainfall amount during a multi-day rainfall event, and second, the cumulative rainfall over the past few day(s). For each category, the antecedent soil moisture recorded at three levels of soil depth, the top 5 cm by satellite-based technologies as well as the top 50 cm and 2 m by modelling approaches, was included in the statistical models to assess its potential for landslide hazard assessment and warning capabilities. The results reveal the cumulative 3 d rainfall
Landslides are one of the most prevalent hazards in mountainous regions of
the world, associated with high rates of fatalities, injuries and economic loss globally (Froude and Petley, 2018; Haque et al., 2016; Kirschbaum et al., 2015; Petley, 2012). According to a recent estimate (Froude and Petley, 2018), precipitation-induced landslides were responsible for a global total of
Rwanda is an evergreen landlocked country geographically located between
1–3
Location of Rwanda in Africa, elevation, spatial and temporal distribution of hazardous landslides with light to dark red dots indicating old to new landslides recorded from 2007 to 2019, landslide representative rain gauges, rainfall distributions indicated by isohyets (sky blue lines) and precipitation footprint of the 5 km buffer around each landslide.
Geomorphology of Rwanda, landslide representative AWSs (automated weather stations) with soil moisture sensors, landslides in red dots and 5 km buffer zones indicating the research area of interest (ROI) for areal soil moisture acquisition.
Wflow model catchments (Kivu, upper Nyabarongo and Mukungwa) and hydrogeology, landslides in red dots and 5 km buffers indicating the ROIs for areal soil moisture acquisition from the Wflow model, and AWSs with soil moisture sensors for comparative performance evaluation of the Wflow-modelled soil moisture.
The inventory for this study contains landslides recorded from 2007 to 2019. It was accessed from a previous study in Rwanda (Uwihirwe et al., 2022) and was further extended and updated through compilation of additional rainfall-induced landslides as reported from local newspapers, blogs and government technical reports. This landslide inventory was compiled with respect to the methodology adopted by Kirschbaum et al. (2010, 2015) and Monsieurs et al. (2018b). Between 2007 and 2019, the inventory includes 55 accurately dated landslides, 32 of which are located in the catchments modelled for this study (Kivu, upper Nyabarongo and Mukungwa) (Fig. 3). However, it is important to note that this inventory is likely to miss the non-hazardous landslides, which are less reported upon than hazardous landslides that led to fatalities/injuries and considerable damages. The inventory provides the location of each recorded landslide but with a varying spatial accuracy of 5 to 25 km depending on the smallest administrative unit recorded by the landslide event reporters. Therefore, a buffer zone of 5 km, equivalent to the frequently recorded accuracy, was used around each landslide (Fig. 1) to support the choice of the landslide representative rain gauge. The same areal buffer was used as a footprint to avail the areal satellite precipitation and soil moisture as detailed in Sect. 3.2 and 3.3.
We accessed daily precipitation data from 19 rain gauges operated by the Rwanda Meteorology Agency. These rain gauges were selected based on their
location within the defined buffer of 5 km around each landslide location (Fig. 1). Once two or more rain gauges fall within the same buffer zone, the gauges are weighted (Melillo et al., 2018) to select the most representative rain gauge following Eq. (1):
With the gauge-based precipitation data as a reference, we assessed the performance of seven satellite precipitation products summarised in Table 1. These satellite precipitation products were preliminarily selected for analysis based on the criteria that their dataset (i) at least partially overlap with the landslide inventory period (2007–2019), (ii) have at least daily temporal resolution, and (iii) be available on the Google Earth Engine (GEE).
Pre-selected precipitation products and short descriptions.
Among the pre-selected satellite products, we have chosen the most suitable
product for landslide hazard assessment in Rwanda based on the relative
comparison with gauge-based precipitation. This was achieved using a number
of statistical approaches that include (i) the use of statistical metrics of goodness of fit, (ii) rainfall frequency indicators, and (iii) intensity
comparisons. The statistical metrics of goodness of fit include the root mean square error (RMSE), Pearson correlation (CC), and long-term relative bias (RB) computed with Eqs. (2) to (4):
The rainfall frequency indicators specify the frequency of rainy days based
on the predefined threshold indices (Joshi et al., 2014; Tank et al., 2009). We used five rainfall threshold indices that reflect the number of rainy days with
In situ soil moisture data, collected from the automatic weather stations (AWSs) equipped with soil moisture sensors, were accessed from the Rwanda Meteorological Agency for six AWSs as shown in Fig. 2. The AWSs recorded the soil moisture at 20 cm depth with a temporal resolution of 5–10 min from July 2018 to December 2019. Because the analysis focuses on a daily timescale, we computed and used the daily average soil moisture time series recorded from July 2018 to December 2019. The in situ AWS soil moisture data were used as a benchmark to comparatively get an insight into the quality of other sources of soil moisture products that include satellite- and model-derived soil moisture estimates described in Sect. 3.3.2 and 3.3.3.
We used a satellite-derived near-surface soil moisture product provided by Planet, formerly VanderSat (VdS) (
We also used the soil moisture derived from Wflow, open-source software developed by the Deltares OpenStreams project (Schellekens, 2021; Schellekens et al., 2019). The Wflow-distributed hydrological model platform currently contains 13 models (Schellekens, 2021) that include the wflow_sbm
model. The models consist of a set of python programs that are run on the
PCRaster python framework to perform hydrological simulations (Karssenberg, 2014; Karssenberg et al., 2010). The Wflow_sbm uses the conceptual bucket model approach to derive the hydrological variables of interest (Imhoff et al., 2020; Schellekens et al., 2019). With Wflow_sbm, the soil is considered a bucket with a depth (
The unsaturated store
The daily rainfall data from the satellite product were used to define the
landslide meteorological triggering conditions. We used two categories of
landslide-triggering conditions. The first category defined a landslide trigger as the maximum probable rainfall event (MPRE), during which or after its end one or more landslides occurred. The MPREs were defined as individual periods of rainy days interrupted by dry periods of at least 2 d. Given the constraint of overestimation of the number of rainy days with 0–10 mm by satellites, a rainy day was objectively referred to as a day with
The normalisation of soil water content
The tested hydrological conditions therefore include the near-surface soil moisture Se
The landslide test variables which include the predisposing hydrological
conditions Se
Since the ROC curve only indicates all possible thresholds and their relative balance between TPRs and FPRs, one is free to choose the optimum threshold depending on whether to maximise the TPR or minimise the FPR. However, according to Postance and Hillier (2017), the optimum threshold is the one that maximises the TPR while minimising the FPR. Therefore, that optimum threshold levels above which landslides are highly likely to occur have been defined using two statistical metrics, i.e. the maximum true skill statistic (TSS) and the minimum radial distance (Rad). The TSS is expressed as a balance between the TPR and FPR as indicated in Eq. (12):
The optimum thresholds defined based on the maximum TSS and/or minimum Rad were plotted in a 1D threshold space, here referred to as the single-variable threshold line, beyond which the probability of landslides is high. We also
followed the cause–trigger concept (Bogaard and Greco, 2018) that reflects the hydro-meteorological thresholds and hypothetically plotted the optimum thresholds of the landslide-predisposing hydrological variables, i.e. the antecedent soil moisture on the
The suitability of satellite precipitation products in the study region was
assessed using three statistical indicators as summarised in Table 2 and Table 3 and illustrated in Fig. 4. From the statistical measures of fits (RMSE, CC, RB), it is generally observed that IMERG is consistently more suitable, while ERA5 was found to be the least suitable product as compared to other satellite precipitation products. The evaluation based on frequency indicators is summarised in Table 3. These indicators give an overview of whether a given satellite product would overestimate or underestimate the
observed gauge precipitation based on the predefined threshold indices.
The Integrated Multi-Satellite Retrievals for Global Precipitation Measurement (GPM_IMERG) displays the highest skill in estimating all ranges of rainfall from heavy to extremely heavy rainy days as recorded by the on-site gauges. CHIRPS and TRMM 3B42 v7 provide good estimates of precipitation, with a quite similar number of rainy days (
Intensity comparison between satellite-based and gauge-based precipitation based on the cumulative 30 d rainfall.
Performance of satellite precipitation products based on statistical metrics.
Performance of satellite precipitation products based on rainfall frequency indicators.
Event-based variable thresholds and prediction capabilities.
The suitability of satellite products was also assessed using an intensity comparison indicated by the density of the scatter points around the
Satellite- and model-derived information and landslide activities:
Figure 5 shows the GPM-based IMERG precipitation spatially averaged over all landslide precipitation footprints and over the modelled catchments. It also shows the temporal dynamic of the satellite-derived soil moisture Se
Receiver operating characteristic (ROC) curves, area under the curves (AUC) and optimum landslide thresholds defined by the true skill statistic TSS (square-shaped marker) and radial distance Rad (cycle-shaped marker) using
Figure 6 and Table 4 show the derived landslide meteorological and hydrological thresholds and their predictive capabilities in terms of TPR and FPR. The discriminatory power of each of the tested variables was evaluated with a ROC curve and the AUC statistical metrics as shown in Fig. 6. Among the tested landslide-triggering meteorological variables
Figure 6c and d indicate that the wetness state of soil prior to the
cumulative rainfall
Landslide hydro-meteorological thresholds and prediction capabilities:
The threshold definition metrics, TSS and Rad, resulted in quite comparable landslide thresholds as summarised in Table 4. It was noticed that the
defined satellite precipitation thresholds are similar to the ones defined
using gauge-based precipitation. For example, the optimum landslide threshold event rainfall volume
Landslide hydro-meteorological thresholds and prediction
capabilities:
With respect to the high rate of false positives resulting from the single-variable thresholds, we have tested whether the incorporation of antecedent
soil moisture information into the rainfall-triggering conditions improves the landslide prediction capability. The optimum single-variable hydrological and meteorological thresholds have therefore been combined into
hydro-meteorological thresholds following the cause–trigger concept in a bilinear framework as shown in Figs. 7 and 8. Figure 7 illustrates the
first category of landslide hydro-meteorological thresholds defined based on
the maximum possible rainfall event
In contrast to the classical precipitation thresholds, the combination of
hydro-meteorological thresholds in a bilinear framework provides an improvement in terms of reduced rate of false alarms by about 30 %
(Se
The IET and the timescale of the rainfall events
Cumulative rainfall-based variable thresholds and prediction capabilities.
This study reveals the high capability of the NASA GPM-based IMERG product to reproduce rainfall patterns which are consistent with the gauge-based precipitation and thus more suitable for landslide hazard assessment thresholds than sparsely distributed rain gauges in Rwanda. However, this research also points out that the IMERG satellite-based product overestimates the number of rainy days whose daily rainfall is between 0 and 10 mm, and thus the mean annual totals. This may lead not only to differences between satellite- and gauge-based landslide thresholds defined under the same locations, but also to the statistical bias, especially when probabilistic methods are used for landslide threshold definition. To address this constraint and to be able to exploit the usefulness of IMERG precipitation in landslide hazard assessment thresholds, we objectively used 10 mm as a threshold to define a rainy day for IMERG precipitation data. This threshold was defined based on the frequency indicator metric adopted as one of the techniques of bias evaluation between ground and satellite-based rainfall. For gauge-based rainfall, 2 mm is generally considered a threshold to define a rainy day and has been defined based on the mean daily potential evaporation (Marino et al., 2020; Peres et al., 2018).
Although the threshold definition of a rainy day (10 mm) may have led to the
omission of some rainfall information, thus shortening the event duration
It is agreed that the consequences of offering FPRs are less harmful in the short term than missed alarms (FNRs), which implies that the best threshold should maximise the TPRs while minimising the FNRs. However, the thresholds in Fig. 7b and e are classical thresholds
The poor performance of the rainfall event-based thresholds concept is due to uncertainties from multiple sources. We hypothetically used the rainfall events as landslide-triggering conditions, defined as individual periods of continuous rain interrupted by at least two dry day periods referred to as the minimum IET. Nevertheless, this definition needs further exploration to be standardised to avoid uncertainties. According to Adams et al. (1987) and Hong et al. (2017), the IET is defined as the minimum period of time that separates two consecutive rainfall events and is considered the period for which the effects of the antecedent soil moisture or precipitation index may last. This is to say that the antecedent soil moisture and/or antecedent precipitation index have no significant effect on landslide initiations once the rainfall events and IETs are well defined. However, the IET, the period during which the effect of antecedent soil moisture becomes null, depends on a number of site-specific factors (soil properties, land use/land cover, potential evaporation, etc.) and is thus difficult to be standardised. Another drawback associated with the use of rainfall event concept may be linked to the transient timescales of the triggering events that bring about difficulties to fix the appropriate time to give an alert or an early landslide warning to the threatened community.
Looking at the constraints associated with IET, rainy day and rainfall event definitions, we explored the shorter-scaled triggering rainfall conditions that include the cumulated rainfall with constant durations of 1, 2 and 3 d (
Among the tested pre-wetting conditions, the incorporation of the antecedent
soil moisture modelled at the root zone Se
The adopted bilinear threshold framework, indicating the distribution of data points in a 2D space, reflects the relationship between the landslide causal and triggering conditions despite other linked constraints and limitations (Conrad et al., 2021). We objectively used the bilinear threshold framework because the majority of positive classes were clustered in the upper-right corner of the 2D threshold space. Although this format proved to be suitable for landslide hydro-meteorological threshold definition (Mirus et al., 2018; Thomas et al., 2019; Uwihirwe et al., 2020, 2022), other formats could also be useful, depending on the distribution of the positive classes in the 2D space. The adopted bilinear framework is in line with the goal of the hydro-meteorological cause–trigger-based threshold concept that prioritises the minimisation of false alarms while at least keeping unchanged the rate of true alarms. Additionally, in some cases, single-variable thresholds lead to high prediction capabilities in terms of elevated rates of true alarms and with quite low rates of false alarms and could be adopted especially for hydrologically based thresholds that consider the long-term wetting process of the soil until the landslide day.
Regardless of the good performance of soil moisture as a landslide hydro-meteorological threshold, the incorporation of the pre-wetting state of
soil into landslide hazard assessment thresholds using groundwater levels,
Ideally, one would have a landslide inventory of about 200 landslide events in order to have a precise estimation of threshold parameters (Peres and Cancelliere, 2021). However, the landslide inventory used for this study accounts for only 32 hazardous landslides. Although the reliance on this limited sample size is likely to lead to a bias towards the larger landslide events and those with impacts on society, this landslide inventory is the most comprehensive one currently available in the study area.
This research aimed to evaluate the potential of satellite-based measurements of precipitation and soil moisture as well as hydrological model-derived soil moisture information for landslide initiation thresholds in Rwanda. The GPM-based IMERG rainfall product was found to be a good spatially distributed source of rainfall data for landslide hazard assessment, especially in data-scarce areas like Rwanda. The satellite- and model-derived soil moisture time series broadly reproduce the most important trends of the in situ soil moisture. Regardless of different depths of data records and slight overestimation of soil moisture by satellite- and model-derived techniques, it was concluded that they follow the in situ observed temporal variation and are thus potentially useful for landslide initiation threshold definition. The purpose of incorporating the antecedent soil moisture into landslide hazard assessment was to account for the physical effect of the pre-wetness state of soil, responsible for the predisposal of the slopes to near failure, prior to the landslide-triggering conditions. Two categories of landslide-triggering conditions have been considered to assess the potential value of including the antecedent soil moisture information. The category that considers the cumulative 3 d rainfall was the most impactful and thus was more useful for landslide hazard assessment in Rwanda rather than the rainfall-event-based trigger. Although the area under the curve (AUC
Selected examples of satellite- and model-derived soil moisture compared to in situ recorded soil moisture at 20 cm soil depth (AWS):
The landslide inventory used for this research can be accessed at
JU collected the in situ data, conducted the statistical analysis and conceptualised and prepared the manuscript storyline. ARP collected and prepared data from satellite precipitation products, organised the soil moisture data and conducted some statistical analysis. HW and JS prepared and made available the high-resolution satellite-based soil moisture data. FSW actively participated in hydrological modelling of soil moisture and greatly contributed to the correction and perfection of the manuscript. MS corrected the manuscript storyline, shaped the discussion and contributed to the perfection of the manuscript. TAB initiated the research idea, proposed the research approaches, created the research network and collaboration, verified the data preparation and statistical analysis, shaped the manuscript storyline and contributed to the perfection of the manuscript.
The contact author has declared that none of the authors has any competing interests.
Publisher's note: Copernicus Publications remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This article is part of the special issue “Hydro-meteorological extremes and hazards: vulnerability, risk, impacts, and mitigation”. It is a result of the European Geosciences Union General Assembly 2022, Vienna, Austria, 23–27 May 2022.
We are so thankful to Elise Monsieur, Arthur Depicker and Olivier Dewitte for sharing part of the Landslide Inventory for Rwanda as part of the central section of the Western branch of the East African Rift (LIWEAR) project. We are thankful to the Rwanda Meteorological Agency for offering access to the in situ meteorological and hydrological datasets used for validation of other sources of data.
This paper was edited by Francesco Marra and reviewed by two anonymous referees.