Keywords

1 Introduction

In tropical regions, rainfall is one of the main triggering factors of landslides: Monitoring, early warning, and mitigation; all these three strategies are equally crucial for landslide disaster risk reduction. The current study is a case study from Aranyaka, Sri Lanka, where a devastating landslide occurred in the first monsoon of 2016, May 17th, causing severe human and economic loss. Even though the vulnerable area was already identified as a landslide high hazardous zone, adequate local early warning and evacuation advisories could not be performed timely. Sri Lanka receives high-intensity rainfall as a tropical island during two major monsoons each year: the first monsoon (South and Southwestern region) in May to September and the second monsoon (North and Northeastern region) in December to February. During these seasons, catastrophic landslides have occurred periodically, and records show an increasing tendency of events and losses.

Accurate rainfall prediction is necessary for reliable landslide early warning; therefore, high-resolution numerical weather simulations are important (Kamiya et al. 2019; Onishi et al. 2019). Numerical weather simulation permits mathematical modeling in global, regional, or local scale weather predictions for long-term and short-term forecasts based on current weather. Numerical weather simulation systems which are currently used for rainfall prediction in Sri Lanka, such as ECMWF (European Centre for Medium-Range Weather Forecasts) providing 11 km resolution rainfall data and WRF (Weather Research and Forecasting Model) running with 9 km- and 3 km-resolution simulations, are still not strong enough in providing finer resolution rainfall predictions in landslide early warning perspective. No local weather simulation model is currently operating for rainfall forecasting in Sri Lanka. There need to be more local resources to ensure the safety of vulnerable communities on landslides.

Multi-Scale Simulator for the Geoenvironment (MSSG) (Takahashi et al. 2006, 2013), developed by the Tokyo Institute of Technology, Japan Agency for Marine-Earth Science and Technology and Waseda University, is a coupled non-hydrostatic atmosphere-ocean-land global circulation model that permits seamless simulations between weather and climate, serving high-resolution simulations. This MSSG can provide local rainfall predictions that can meet the need for the landslide early warning system in Sri Lanka. MSSG can be run as a multiscale numerical weather simulation model (Fig. 1). It allows to run global simulations with O(1–10 km) horizontal resolutions and regional simulations with O(100 m)–O(1 km) horizontal resolutions. MSSG is also developed to run local scale (urban scale) simulations with O(1 m) horizontal resolution resolving buildings and considering anthropogenic heat releases. In terms of the simulation scales, the high efficiency and reliability of MSSG have been confirmed in all scales using powerful supercomputers (Onishi et al. 2022). In global scale simulations, MSSG uses the Yin-Yang grid system (Kageyama and Sato 2004; Baba et al. 2010), which allows for overcoming the effect of polar singularity and lead to high computational efficiency. In the regional scale simulations, MSSG uses a conventional latitude-longitude system to allow using robust high-order differential schemes.

Fig. 1
3 illustrations of atmospheric simulations on a global scale to a horizontal scale of 1 to 10 kilometers, a regional scale to a horizontal scale of 100 meters to 1 kilometer, and a local scale to a horizontal scale of 1 meter to 100 meters. The images form a 1 way or 2-way nesting scheme.

Scales of atmospheric simulations. The nesting domain system is used in MSSG multiscale simulations. (a) Global simulations use Yin-Yang grid system, (b) regional simulations use the conventional longitude-latitude orthogonal grid system, (c) local (urban) simulations allow building and residential topography resolved

Full size image

High-resolution simulations are reliable, but at the same time, they require high computational costs. In case we develop an operational rainfall prediction system, the balance between reliability and computational cost. For example, an operational prediction system would be useless if it took 24 h for 24-h prediction. We usually need to compromise on reliability if we use an ordinary computer with limited performance unlike supercomputers. However, the current machine learning technology would relax the compromise in reliability. The deep-learning-based super-resolution technology would help increase the resolution, leading to better reliability at a low computational cost.

We assess the reliability of MSSG high-resolution simulations in comparison with the local observation data in Sri Lanka. We also discuss the potential of using machine-learning-based super-resolution in operational real-time rainfall predictions for landslide early warning systems.

2 Study Area

Figure 2 shows the map of Sri Lanka, centering the focus area of the present study, Aranayaka; more specifically, Samasarakanda in the Aranayaka Divisional Secretariate in the Kegalle District of Sri Lanka, where the devastating landslide occurred in 2016. The area is in the country’s central highlands and experiences high yearly rainfall during the southwestern monsoon yearly. An automated rain gauge recorded 426 mm cumulative for 58 h continued rainfall starting from 15th May 2016, which triggered the rapid landslide, which traveled downslope about 2 km air distance causing about 150 human losses.

Fig. 2
A map of an island surrounded by a 512 by 512 kilometers square border. A 128 by 128 kilometer margin is on the island. The center point is Aranayake which has a smaller square margin of 32 by 32 kilometers.

The Google Earth picture of Sri Lanka, centering the study point, Aranayaka. The simulation domains utilized in this study are overlaid

Full size image

3 Methodology

Three computational domains (parent, child and grandchild domains) were selected. The Center of the domains was set at Aranayaka, latitude 7.1541°N, and Longitude 80.4296°E for all the three domains. The parent domain size was selected to cover most of Sri Lanka. The landslide occurred on the slope of a mountain with the half-width of about 2 km. The horizontal resolution of the grandchild domain was set to 500 m considering the reasonable horizontal spacing on the scale of landslides and limitations due to the performance using a personal computer. Table 1 summarizes the horizontal settings, i.e., longitudinal and latitudinal settings. For the vertical coordinate, 32 grids with irregular intervals were set for all three nesting domains for the height of 30 km. These domain and grid sizes were determined considering the perspective and computer performance.

Table 1 Settings of the computational domains
Full size table

Numerical weather models compute the advection, i.e., transport of various tracers, including water substances such as water vapor, cloud water, rain, snow, cloud ice, and graupel (Miltenberger et al. 2015). For accurate computation in advection, the fifth Wicker-Skamarock (W-S) scheme was employed (Wicker and Skamarock 2002).

The six-category bulk cloud microphysics (MSSG-Bulk) (Onishi and Takahashi 2012) was employed to explicitly calculate the cloud dynamics.

4 Results and Discussion

4.1 Five-Day Rainfall Simulation for the Aranayaka Landslide on May 2016

MSSG with the 8 km–2 km-500 m nesting grid system was performed for the rainfall from 00UTC on May 14 to 00UTC on May 19, 2016. Figure 3 shows the rainfall results from the MSSG simulation compared with the observed rainfall recorded in the automated rain gauge of the National Building Research Organization, Sri Lanka, located at Aranayaka. Table 2 summarizes the values of cumulative rainfall and maximum rainfall obtained from observation, GSMAP (the Global Satellite Mapping of Precipitation), and MSSG. As in the table, cumulative rainfall of 335.7 mm over 5 days was predicted through the simulation, while the observatory recorded 446.3 mm accumulation for the considered period. The maximum hourly average rainfall resulted as 22.7 mm/h through simulation, while the observatory recorded 26.8 mm/h. From the GSMAP data for the above period, cumulative rainfall of 192.0 mm and maximum hourly rainfall of 12.7 mm/h were recorded. When considering this 5-day rainfall event as a whole, the bell-shaped intensity distribution of the long-lasting continuous rainfall is compatible up to a reasonable extent between the observatory record and the MSSG simulation. Compared with the GSMAP, which provides global hourly rainfall at 0.1 × 0.1-degree resolution, which roughly corresponds to 10 km × 10 km horizontal resolution, based on satellite data, the MSSG simulation with 8 km–2 km-500 m resolutions provided closer values to the observatory record. This result confirms that high-resolution simulations are needed for better local rainfall predictions in mountainous areas.

Fig. 3
A waveform of rainfall intensity versus date and time in U T C. The highest observatory data is between 8 to 10 on May 17, 2016. The data for the M S S g simulation is the highest during 3 p m on May 16, 2016. The data for G S MAP is highest at 4 a m on May 15, 2016. Values are estimated.

Rainfall comparison of MSSG 8 km_2 km_500 m simulation, observatory and Rainfall forecast by GSMAP, for the Aranayaka Landslide event for period of 5 days

Full size image
Table 2 Rainfall accumulation and maximum rainfall rate at Aranayaka from 2016/5/14 00UTC to 2016/5/19 00UTC
Full size table

The MSSG simulations conducted for the Aranayaka landslide event, utilizing the 8 km–2 km-500 m resolution, successfully reconstructed a rainfall event that aligned well with the observations from the observatory. Although there were variations in the timing and intensity of the rainfall peaks, the overall integration of the event yielded good results. It is worth noting that while comparing point data may have certain limitations, the analysis of the horizontal distribution of rainfall generated by the MSSG simulation indicates highly accurate and reliable predictions. This result suggests that even with the potential drawbacks of point data comparison, the simulation’s ability to depict the spatial distribution of rainfall is remarkably precise and dependable.

4.2 Rainfall Simulations for Past Heavy Rain Events on Aranayaka

Secondly, MSSG simulations were conducted to investigate six previous rainfall events with the high recorded rainfall accumulation in the Aranayaka rain gauge. Here we performed two sets of MSSG simulations to investigate the impact of increased resolution: One is with 8 km–2 km-500 m nesting grid system and the other with 8 km–2 km nesting grid system.

The results are summarized in Table 3. The findings revealed that the 8 km–2 km-500 m resolution simulation mainly exhibited higher maximum rainfall intensity and more significant overall rainfall accumulation than the 8 km–2 km resolution. In rare instances, the inverse trend is observed, though.

Table 3 Summary of the simulations performed for high rainfall-accumulation events in the record of Aranayaka automated rain gauge
Full size table

Figure 4 shows the rainfall comparison for event 3 (in Table 3), which resulted in very close values of maximum hourly precipitation and cumulative rainfall during the event. The values of (a) observatory data, (b) MSSG simulation with 8 km–2 km-500 m, (c) MSSG simulation with 8 km–2 km, and (d) GSMAP for maximum hourly average precipitation are, respectively, 16, 13.2, 4.4 and 7.48 mm/h, and for cumulative precipitation, they are, respectively, 114.5, 123.1, 41.1 and 25.81 mm/h. In terms of the timing of major rainfall, MSSG prediction shows the peak at a later time. This time shift can be partly explained by the initial condition of the MSSG simulation. The initial data, created from GSM data by the Japan Meteorological Agency, did not contain cloud substances. This prevents the simulation from producing rainfall in the early time. This initial problem can be solved if we restart the simulation from the previous simulation results. This time, we did not use this solution since we focused on the influence of the simulation resolutions.

Fig. 4
A line graph plots rainfall intensity versus date and time in U T C. Line a peaks at 9 p m, line b peaks at 3 p m, line c peaks at 12 a m, and line d peaks at 4 p m.

Rainfall comparison of (a) the observatory, (b) MSSG 8 km–2 km-500 m simulation, (c) MSSG 8 km–2 km simulation and (d) rainfall data by GSMAP for Event 3, from 18UTC on 31 July to 18UTC on 1 August, 2022

Full size image

The results obtained from the second step of reconstructing high-intensity rainfall events in the past shed light on an important finding: increasing the resolution of the simulations lead to higher values of maximum and cumulative rainfall.

Figure 5 illustrates the increasing accuracy of simulation with increasing resolution. This observation indicates that as the simulation resolution increases, it becomes more capable of capturing the local variations associated with local topography. The relationship between resolution and the ability to accurately represent precipitation patterns highlights the importance of considering finer scales in meteorological simulations to understand better and predict intense rainfall events. Moreover, this study showed that the MSSG simulations were generally successful in simulating the development of isolated or localized cumulonimbus clouds. However, when the size of these cumulonimbus clouds was relatively small or the duration of rainfall was short, reproducing their exact location through the MSSG simulations remained challenging. Nevertheless, from the current perspective, it is essential to note that landslides are typically not triggered by brief, isolated rainfall events but rather by prolonged periods of continuous rainfall or high-intensity rainfall with a significantly longer duration. The accumulation of water in the soil over time ultimately leads to an increase in mass and loss of cohesion, facilitating downslope movement or rapid erosion. Therefore, the challenges in reproducing the exact location of isolated cumulonimbus clouds do not significantly impact the accuracy of rainfall predictions in the context of landslide early warning systems.

Fig. 5
3 heat maps of the Sri Lankan island. Sri Lanka's geographic center is enlarged. The magnified land exhibits different erratic zones in shades.

Regional distribution of rainfall at the time of maximum rainfall through simulation showing the effect of increasing resolution in different domains, (a) 0C_parent domain which the grid size is 8 km × 8 km, (b) 1C_child domain which the grid size is 2 km × 2 km, (c) 2C_grand child domain which the grid size is 500 m × 500 m

Full size image

5 Future Work

5.1 Turbulence Enhancement of Orographic Rainfall

This study used the “typical cloud model”, i.e., MSSG-Bulk, installed in MSSG simulation system. Recent research has proven that the turbulence enhancement of collision growth in orographic settings significantly impacts rainfall enhancement (Falkovich and Pumir 2007; Grabowski and Wang 2013; Onishi et al. 2015). The present MSSG have the “turbulence-aware cloud model”, developed by Seifert and Onishi (2016) and Onishi and Seifert (2016), which is capable of considering turbulence enhancement in orographic rainfall.

According to Seifert and Onishi (2016), and as illustrated in Fig. 6, turbulent enhancement in orographic rainfall can impact conversion rates among different categories, i.e., cloud droplet growth.

Fig. 6
2 illustrations. A. An illustration of droplets from the topographical updraft condensing to form clouds. B. A cyclic representation of cloud water undergoing condensation to form water vapor. The cloud water coalesce to form rain that precipitates. Rain evaporates to form vapors.

(a) Schematic illustration of turbulence enhancement of orographic rain, (b) interactions and processes among categories of warm cloud. Pauto represent amount of increase in raindrops due to collision of small water droplets and merger into large water droplets, and Paccr implies the amount of increase in raindrops due to large water droplets collecting and merging small water droplets. This two coalescence factors Pauto and Paccr are enhance due to turbulence

Full size image

Autoconversion, Pauto, represents the conversion rate from the cloud-water category to the rain-water category due to collisions among cloud droplets. Accretion, Paccr, represents the conversions rate from cloud water to rain-water category due to collections of cloud droplets by rain drop.

The turbulence enhancement in drop collisions enlarges the two conversion rates, Pauto and Paccr. The present MSSG simulations did not use turbulence-aware cloud models and they are, therefore, not capable of considering turbulence enhancement. The impact of turbulence enhancement in past rainfall events should be assessed as a next step.

5.2 Super-Resolution Simulations

Downscaling techniques, including machine learning models such as artificial neural networks and support vector machines, are employed to enhance the spatial resolution of low-resolution models.

High-resolution (HR) numerical simulations offer superior predictive capabilities to low-resolution (LR) simulations but come at a higher computational cost. The super-resolution (SR) simulation system (Onishi et al. 2019) employs a two-step approach to reduce the computational cost for HR prediction information. Firstly, an LR simulation is conducted, and then the resulting LR prediction images are mapped to HR predictions using an SR method. This combination enables the system to provide HR predictions as reliable as those obtained from HR simulations while at a much smaller computational cost.

A recently introduced super-resolution (SR) neural network, as proposed by Yasuda et al. (2022), demonstrates its applicability in generating rainfall prediction maps. The SR network can utilize multiple input channels, such as a high-resolution topography map, to enhance the accuracy of the resulting high-resolution (HR) rainfall map and to widen the generality of the network. The SR neural network is trained using HR and low-resolution (LR) simulation results during the training phase. HR simulations, which are computationally intensive and performed on supercomputers, contribute to generating training data sets. In contrast, during operational use, LR simulations require significantly lower computational resources. This SR system holds promise for providing real-time prediction services to local communities needing more computational capabilities.

6 Conclusions

The Multi-Scale Simulator for the Geoenvironment (MSSG) offers a promising tool for enhancing regional simulations with high resolution. The MSSG simulation with an 8 km–2 km-500 m nesting grid system has demonstrated reasonably accurate rainfall predictions, particularly for investigating local orographic rainfall. Occasionally, less accurate rainfall predictions can be attributed to the challenge of reproducing the exact location of short-lived, isolated cumulonimbus clouds.

The MSSG provides more reliable rainfall data as the resolution increases. This study has revealed that by increasing the resolution from 8 km–2 km to 8 km–2 km-500 m, the cumulative rainfall and maximum hourly-average rainfall for specific events exhibit higher values, which align closer to the observed data. This finding establishes the reliability of the MSSG, paving the way for the confident implementation of fine-resolution simulations.

The real-time operation using the MSSG with an 8 km–2 km-500 m nesting grid system would require expensive computers. The deep-neural-network super-resolution technology would help to reduce the computational cost to produce high-resolution reliable predictions for orographic rainfall. This technology combination positions the MSSG as an effective tool for rainfall forecasting and early warning of landslides.