Landslides are pervasive in mountainous areas and adversely affect human life and infrastructure (Chester 1995; Peel et al., 2007). Key factors leading to landslides can be broadly classified into two categories: the dynamic triggering mechanisms like extreme rainfall and local soil moisture conditions and prevalent ground conditions (Guzzetti et al., 2020). It is critical to develop near real-time and early warning systems for predicting impending events. These systems could provide, emergency managers and planners, advanced situational awareness to save lives and property. Predicting when and where landslides may occur and issuing warnings, accordingly, is an evolving area of research. Landslide models can either be designed to function as near real-time or forecasting systems. Literature refers to predictive systems using several different terms, such as landslide models, forecast models, warning models, and warning systems. A forecast model provides estimate of the future state of a natural system obtained with a numerical model (Ramage, 2022). A “warning model” is a framework to issue landslide advisories and can incorporate one or more landside models and advisory protocols that are used to issue advisories. Comparatively, a “warning system” is a physical implementation of a warning model with one or more landslides forecast models (Guzzetti et al., 2020).

Landslide Early Warning Systems (LEWSs) are usually developed locally and are designed for specific uses, such as civil protection, hill slope monitoring, and regional situational awareness. According to literature, a total of five nations, 13 regions and the metropolitan areas of Chittagong, Hong Kong, and Seattle benefit from LEWSs (Guzzetti et al., 2020). An example of a near-real time system is Rio de Janeiro Brazil’s Alerta Rio system, which fuses information from susceptibility maps with the rain-gauge measurements, updated every 5 min (Calvello et al., 2015). Some examples of LEWSs that leverage numerical weather model Quantitative Precipitation Forecasts (QPF) include one Ahmed et al., 2018 proposed (operated on voluntary basis) for Chittagong Metropolitan Area, Bangladesh. This LEWS exploits empirical rainfall thresholds, daily QPF, and a statistically based landslide susceptibility zonation for preparing landslide forecasts. The Civil Protection Department of Sicily, southern Italy is operating a Hydrohazards Early Warning System (HEWS: flood and shallow landslide warning system) based on the combined use of rainfall thresholds, soil moisture modelling and QPF (Brigandì et al., 2017). Similarly, the regional government of Piedmont, northern Italy, is operating a regional LEWS which comprises three complementary landslide forecasting systems, namely, the DEFENSE (Tiranti et al., 2014), SMART (Tiranti and Rabuffetti, 2010) and TRAPS (Tiranti et al., 2013) systems. LEWS based on the comparison between hourly rainfall measures, rainfall forecasts up to +48 h, and 3D rainfall thresholds (using intensity, duration, and antecedent rainfall as rainfall parameters) is operated in the Emilia Romagna Region, Italy (Rosi et al., 2021). The Japanese Meteorological Department’s early-warning system is based on hourly rainfall, short-range precipitation forecasts, and soil-moisture index (Singh, 1995; Sugawara et al., 1983; Osanai et al., 2010), and the Norwegian national landslide early warning system uses hydrologic models and web tools to monitor and forecast hydrometeorological conditions that could potentially trigger landslides (Graziella et al., 2015). However, these LEWSs cover a very small percentage of the land susceptible to landslides globally (Nadim et al., 2006). Results compiled by Froude and Petley, (2018) suggest that most of these national and regional LEWSs do not operate where a large majority of fatal landslides occur and the risk of landslides to the population is high.

Kirschbaum et al. (2018) proposed a quasi-global near-real time Landslide Hazard Assessment for Situational Awareness (LHASA) model. LHASA combines satellite-based precipitation estimates from the Global Precipitation Measurement (GPM) mission with a landslide susceptibility map derived from information on slope, geology, road networks, fault zones, and forest loss, primarily from satellite-derived or publicly available data (Kirschbaum et al., 2018; Stanley and Kirschbaum, 2017). Version 2 of LHASA leverages machine learning to produce probabilistic ratings of landslide hazard at 1 × 1 km resolution at quasi-global scale (60°N-60°S) (Kirschbaum et al., 2020; Stanley et al., 2021). In addition to GPM rainfall estimates, LHASA version 2 ingests data on snow mass and soil moisture from the Soil Moisture Active Passive Level 4 (SMAP L4) product (Reichle et al., 2018). It also analyzes the exposure of population and road networks to landslide hazard for each level-2 administrative district (Emberson et al., 2020). Although LHASA version 2 provides near-global information in near-real time (∼4hrs latency), many stakeholders have expressed a preference for forecasting of landslide hazard, which is not addressed solely using satellite observations.

In this study, we test the feasibility of ingesting a global precipitation forecast from the NASA Goddard Earth Observing System Forward processing product (GEOS-FP; Molod et al., 2012; Rienecker et al., 2008) into the LHASA framework to provide probabilistic landslide estimates for one to 3 days in the future. As landslide hazards can be triggered in relatively short time, the global landslide forecast system could serve as a tool in conjunction with other situational awareness products for an impending major rainfall event (e.g., tropical storms). This represents a new effort in global landslide forecasting and a first attempt to provide a global predictive view of landslide hazard, which can be particularly useful in areas without ground-based systems or active monitoring programs. This work introduces the global LHASA-Forecast (LHASA-F) framework and evaluates its performance relative to the LHASA near-real time (LHASA-NRT) model and observed landslides. The overarching goal of this work is to determine the feasibility of incorporating a global forecast precipitation product within the LHASA framework to represent potential landslide hazards into future. This study considers analyses at global, regional, and local scales, with an emphasis on validation studies in multiple geographic and climatological settings using landslide inventories.

Landslide forecasting is a challenging task due to several complex phenomena which may contribute toward landslide occurrence. This is further complicated by the inherent uncertainties in the data used for developing landslide models. As in any modeling technique, simplifying assumptions are often employed to reduce the complexity in formulating underlying models. The goal of this study is to evaluate the feasibility of using a global precipitation forecast within the LHASA global landslide modeling framework to better anticipate future landslide hazard. Landslide inventories from representative areas are used as an independent reference to evaluate the performance of both products at a regional scale. Comparing LHASA-F to LHASA-NRT provides a first step in characterizing the regional differences in landslide predictions using the outlined framework.

LHASA-F evaluation framework designed in this study is based on following assumptions: Event-based landslide inventories are complete and do not miss any landslides triggered by the event; LHASA-NRT is a reasonable comparator, in the absence of an alternative global NRT hazard model; p99 calculations are not impacted by different computation time spans used for GEOS-Forecast (2018–2021) and IMERG-Early (2000–2018); substituting rescaled precipitation derived from GEOS-FP in model trained with rescaled precipitation derived from IMERG-Early alleviates the model dependence of raw precipitation. Some of these assumptions are driven by the limited historical global forecast availability. However, as previously stated, all landslide forecasting models have some assumptions (Guzzetti et al., 2020). Despite these limitations, the findings of this study demonstrate promising performance at the global scale in the analysis window and represents a step towards improved predictive capability for global landslides.

Khan et al., 2021 reported that the GEOS-FP model-based precipitation forecast demonstrates coherence with the near-real time satellite estimated (IMERG Early) for tropical storms and extreme precipitation (∼> 100 mm) in general, for regions with landslide susceptibility. As rainfall-triggered landslides are mostly caused by extreme rainfall conditions, the GEOS-FP product was deemed fit to be ingested in the new global LHASA framework for forecasting landslides. LHASA-NRT and LHASA-F exhibit high coherence for landslide points, with both associating higher probabilities (∼>0.70) with major events (Tropical storms), as demonstrated by the scatter plots in Figures 10A–C.

Although LHASA-F generates slightly lower landside probabilities (Figure 5) than LHASA-NRT, these are relative probabilities and the differences may not be significant from operational perspective, especially for extreme events. Results from the case studies presented in this work indicate that the performance of the forecast model varies with specific storm and its geographical location. The accuracy of LHASA-F improves as the forecast time is closer to the prediction time. The reduced performance of the forecast model for 3-days lead time observed in case studies is attributable to uncertainties associated with forecasted precipitation (Sikder and Hossain, 2019). Factors such as initial meteorological conditions, data assimilation methods, and the approximations required to represent physical processes, all contribute towards the forecast skill with lead time. Another reason for degraded performance of LHASA-F with longer lead time could be linked to antecedent conditions used in the model training framework (trained on LHASA-NRT dynamic variables). While the antecedent conditions for Forecast 1-day are derived from IMERG-NRT, those for Forecast 3-days are based on the forecast from the last 2 days (refer to Table 3 for details). However, the global scale mean difference maps between LHASA-NRT and LHASA-F 2-days and 3-days products reveal overall similar trends as shown in Figure 5 for LHASA-F 1-day, with LHASA-F 2- and 3-days showing lower probability values across the globe, and higher values (∼0.03–0.05) in some parts of western South America (SM6 and SM7). The spatial tendency of the LHASA-F system to provide high probability estimates is assessed using a threshold of 0.9. SM3 shows the global map with a count for, each 1 × 1 km grid box, of the number of days when LHASA-F estimated a landslide probability of >0.9, when LHASA-NRT does not. The count ranges from 1 to 5 days out of 365 days of testing period. LHASA-F estimates are lower than high threshold (0.9) most of the days across the globe (shown with white color inland) except for few grid boxes in Jamaica and El Salvador, parts of New Guinea, and north central coast Vietnam.

There are limited global landslide inventories for the relatively short historical GEOS-Forecast record, which did not allow development of a model solely based on the GEOS-Forecast. This problem is addressed by employing rescaled precipitation in both the LHASA-NRT and LHASA-F frameworks. Three dynamic variables, rescaled precipitation, soil moisture, and snow mass from GEOS-FP are used for the validation studies in the model trained with rescaled precipitation from IMERG-NRT, to generate LHASA-F up to 3-days lead time at the global scale. Depending on the availability of a long and consistent precipitation forecast data, and global landslide inventories, future work could address these limitations by training models for each forecast lead-time and combination of inputs.

We find the following overarching results when comparing LHASA-F to LHASA-NRT and available event-based inventories:

• The rescaling of the GEOS-FP precipitation product is a critical step in incorporating the forecasted precipitation data within LHASA-F, though when compared directly, probability values for LHASA-F are low relative to the LHASA-NRT.

Periodic assessment of the landslide forecast system and user feedback is vital for its operational success and utility. This will require assessment of the system by stakeholders using more diverse regional landslide inventories. We envision the availability of such inventories through global initiatives such as LandAware (Calvello et al., 2020) will be crucial for the generalization as well as the advancement of the global landslide forecasting efforts.

Furthermore, updates to IMERG and GEOS may improve the predictive capability of LHASA model and could be a path for future study. LHASA framework is designed with flexibility for adapting to the future updates to the products.

LHASA-F could be used to issue landslide advisories in conjunction with other situational awareness tools. Our analysis at administrative district level 2 shows that landslide forecast outputs may be mapped differently depending on the application. Results of this system are promising, and we continue to engage with stakeholders to support the continual development, localized adaptation, and implementation of the landslide forecast framework. We plan to deploy this model routinely and to make it publicly available under the LHASA framework.

LHASA-F’s global coverage will be valuable in providing crucial information in data-sparse, ungauged, or large-scale catchments to provide broader situational awareness of potential landslide hazard. It is not intended to provide local scale warnings or to supersede local and regional systems that have been calibrated for the specific environment. The LHASA-based hazard estimates provide the likely location of landslide events and inform decision-making about disaster preparedness and response. Additional modules, incorporated into LHASA, such as estimates of population and infrastructure exposure to landslide hazard (Emberson et al., 2020), could be integrated into the landslide forecast to provide further information on potential landslide impacts as well as hazard estimates. The availability of such multi-temporal landslide hazard and exposure tools will equip stakeholders with complementary resources to aid in disaster response, planning and decision-making.