Real-Time High-Resolution Prediction of Orographic Rainfall for Early Warning of Landslides

Abstract

Heavy rainfall often causes devastating landslides. Early warning based on reliable rainfall prediction can help reduce human and economic damages. This paper describes a recent development of reliable high-resolution prediction of orographic (topographic) rainfall using our next-generation numerical weather prediction model, the Multi-Scale Simulator for the Geoenvironment (MSSG). High-resolution computing is required for reliable rainfall prediction, and the MSSG can run with very high resolutions. Robust cloud microphysics is another key to realizing reliable predictions of orographic clouds, where the atmospheric boundary turbulence can affect. This paper clarifies that in-cloud turbulence can enhance cloud development. The recent cloud microphysics model that can consider turbulence enhancement is newly implemented in the MSSG. The emerging machine-learning technology is also coupled with the MSSG for reliable operational predictions. We show the recent development towards reliable predictions of orographic rainfall for realizing early warning of landslides.

1 Introduction

Most tragic landslides are caused by heavy rainfall. For example, the most tragic landslide disaster for recent Japan happened in Hiroshima in August 2014. A series of landslides following local heavy rain reportedly killed seventy-four people. An automated observatory recorded over 200 mm rainfall accumulation within 3 h. This intense rainfall within such a short period made it difficult to issue an early warning, and unfortunately, the evacuation advisory by Hiroshima prefecture was too late.

Reliable rainfall prediction is undoubtedly needed for early warning. There are several numerical weather prediction models available in the world. Examples are WRF (www.wrf-model.org) in the United States, ASUCA in Japan, and COSMO (www.cosmo-model.org) in Germany. These are called regional models as they target regional- (or meso-) scale weather with O(1 km) horizontal resolution. Such regional models are usually for short-term forecasts. However, it is often valuable for early warning if the chance of heavy rain is predicted more than two days before. The regional models need, for forecast, boundary conditions provided by global models. Global models are generally run with O(10 km) horizontal resolutions and guide long-term forecasts (usually more than 2 days). Due to the coarse horizontal grid of O(10 km), the global models need the aid of cumulus parameterization and are usually not good at local heavy rain predictions.

The recent development of supercomputer systems allows us to run global simulations with O(1 km) horizontal resolutions or even finer resolutions with O(100 m) without the aid of cumulus parameterizations. The Multi-Scale Simulator for the Geoenvironment (MSSG; Takahashi et al. 2013 and references therein) can serve for such simulations.

The following subsection briefly reviews the numerical weather prediction models. We then explain the multi-scale weather prediction model MSSG. Some examples of MSSG simulations for heavy rain events are shown. We finally describe the recent integrated technology of machine-learning and computer simulation technologies for operational real-time prediction systems.

2 Multi-scale Weather Prediction Model—Multi-scale Simulator for the Geoenvironment (MSSG)

2.1 Overview

Two types of models are in use for numerical weather prediction (NWP): global and regional models. Global models generally run with O(10 km) horizontal resolutions and guide long-term forecasts (usually more than two days). Regional models are used for shorter-term forecasts and run with O(1 km) horizontal resolutions. Some meteorological processes are too small to be explicitly included for the global models. Then the so-called parameterization is a way for representing these processes by relating them to grid-scale variables. For example, a typical cumulus cloud has a scale of O(1–10 km), much smaller than the horizontal resolutions of global models. The convection process by cumulus clouds should be parameterized, and the parameterization is called cumulus parameterization, on which most of the global models rely.

The recent development of supercomputer systems allows us to run global simulations with O(1 km) horizontal resolutions without the aid of cumulus parameterizations. The Multi-Scale Simulator for the Geoenvironment (MSSG; Takahashi et al. 2006, 2013 and references therein) can serve for such simulations.

The MSSG is an atmosphere-ocean coupled non-hydrostatic model aimed at seamless simulations from global to local scales (Fig. 1). MSSG consists of atmospheric (MSSG-A) and ocean (MSSG-O) components. MSSG adopts a conventional latitude-longitude grid system for regional simulations and the Yin-Yang grid system (Kageyama and Sato 2004; Baba et al. 2010), consisting of two overlapping latitude-longitude grids, thus avoiding the polar singularity problem, for global simulations. MSSG has been serving a wide range of applications. A global atmosphere-ocean coupled simulation was performed at an 11 km horizontal resolution with a nested region at a 2.7 km horizontal resolution. It successfully showed sea surface cooling induced by a typhoon along the track (Takahashi et al. 2013). High-resolution global typhoon simulations were conducted with 7 km horizontal resolutions to clarify the potential of high resolutions for typhoon track predictions (Nakano et al. 2017). High-resolution regional atmospheric simulations were conducted to investigate the influences of the choice of cloud microphysics scheme and that of in-cloud turbulence on the development of clouds (Onishi et al. 2012, 2015). MSSG-O has been used to investigate the dispersion of radionuclides released from the Fukushima Dai-ichi nuclear power plant with 2 km horizontal resolution (Choi et al. 2013) and the effect of wind on long-term summer water temperature trends in Tokyo Bay, Japan, with 200 m horizontal resolution (Lu et al. 2015). MSSG-A has been applied to building-resolving urban atmosphere simulations with 5 m spatial resolutions to clarify heat environments on streets (Takahashi et al. 2013; Matsuda et al. 2018; Kamiya et al. 2019) (Fig. 1).

Fig. 1

MSSG is designed to be applicable to a global scale, b meso scales and up to c urban scales. The Yin-Yang grid system, which consists of two overlapping latitude–longitude grids indicated in blue and red, is adopted for global simulations

2.2 Dynamical Core of the Atmosphere Component

The dynamical core of MSSG-A is based on the non-hydrostatic equations and predicts the three wind components—air density, water vapor mixing ratio, and pressure.

The third-order Runge–Kutta scheme is used for time integrations. The fast terms relating to acoustic and gravity waves are calculated separately with a shorter time step (Wicker and Skamarock 2002). A fifth-order upwind scheme (Wicker and Skamarock 2002) was chosen for momentum advection and the second-order weighted average flux (WAF) scheme with the Superbee flux limiter (Toro 1989) for scalar advection. For turbulent diffusion, the Mellor-Yamada-Nakanishi-Niino level 2.5 scheme (Nakanishi and Niino 2009) is used. The MSSG-Bulk model (Onishi and Takahashi 2012), which is a six-category bulk cloud microphysics model (see the following subsection for detail), is used for explicit cloud physics, and “Model-Simulation-radiation TRaNsfer code” (MSTRNX; Sekiguchi and Nakajima 2008) is used for calculating longwave and shortwave radiation transfers.

2.3 Cloud Microphysics

The bulk cloud microphysics model in MSSG (MSSG-Bulk model; Onishi and Takahashi 2012) computes the temporal evolutions of mixing ratios of water vapor, cloud water, rain, cloud ice, snow and graupel, and in addition, the number density of cloud ice particles. Thus, the bulk model is a one-moment model for warm rain processes and a partial two-moment model for cold rain processes.

Figure 2 shows the cloud microphysical processes considered in MSSG-Bulk. The prognostic variables are the mixing ratios of water vapor, cloud, rain, cloud ice, snow, graupel, and the number concentration of cloud ice; Q_v, Q_c, Q_r, Q_i, Q_s, Q_g and N_i, respectively.

Fig. 2

Cloud microphysical processes in MSSG-bulk

MSSG also has the warm-bin/cold-bulk hybrid cloud microphysical model named MSSG-Bin (Onishi and Takahashi 2012). In that hybrid model, a spectral bin scheme is used for liquid droplets, while a bulk scheme is used for solid (ice) particles. That is, the expensive but more reliable spectral bin scheme treats the relatively well-understood physics of the liquid phase, and the computationally efficient but less robust bulk scheme is used to treat the poorly understood physics of the ice phase.

3 Rainfall Predictions

3.1 Regional Numerical Weather Prediction

MSSG weather simulations have been applied to several heavy rain events. Examples are the locally heavy rain event in Zoshigaya, Japan, on 4th September 2005 and in Hiroshima, Japan, on 19th and 20th August 2014.

The Zoshigaya heavy rain happened on 4th September 2005, associated with mesoscale convective systems caused severe urban flooding in Suginami ward, in the heart of the Tokyo metropolitan city. An MSSG simulation was performed on 1 km horizontal resolution and 32 vertical layers for the computational domain shown in Fig. 3. The initial data were obtained by interpolating the Grid Point Value (GPV) data at 06:00 UTC on 4th September 2005 provided by Japan Meteorological Business Support Center. Figure 4 shows the surface precipitation at 00:00 on 5th September 2005, when a strong rain band lay east of Tokyo and Kanagawa. The local maximum precipitation was underestimated, but MSSG successfully reproduced the south-north rain band.

Fig. 3

Computed area showing a the orographic elevation and b the land use index

Fig. 4

Surface precipitation at 00:00 JST on 4 September 2005 from a the radar observation (Tokyo District Meteorological Observatory) and b the MSSG simulation result

The Hiroshima heavy rain happened on 19th and 20th August 2014, which led to devastating landslides. The regional MSSG simulation was applied to reproduce this event (Hiruma et al. 2022).

Figure 5a shows the horizontal computational domain, which covers a part of the Chugoku, Shikoku, and Kyushu regions, Japan. The number of grid points was N_λ × N_φ × N_z = 648 × 648 × 207, where subscripts λ, φ, and z denote longitudinal, latitudinal, and altitudinal directions, respectively. The horizontal grid spacing Δx_λ = Δx_φ was set to 500 m both for latitudinal and longitudinal directions. The domain height was 30 km, and a damping layer was laid in the top 1/3 of the domain. The vertical grid spacing Δz was varied from 14 m for the lowest layer to 123 m for the highest layer. The simulation start time was set to 9 pm in Japan Standard Time (JST; 9 h ahead to the coordinated universal time, UCM) on 19th August 2014 and the duration to 8 h.

Fig. 5

a Computational domain. b Horizontal distribution of the 6-h (11 pm–5 am) rainfall accumulation over Hiroshima. c A three-dimensional visualization of the cloud distribution over Hiroshima at 00:10 am viewed from the southeast direction. The dashed line corresponds to the solid line in (b)

Figure 5b shows the 6-h rainfall accumulation over Hiroshima. This figure shows that the MSSG simulation successfully reproduced the line-shaped heavy rainfall area, with a length of approximately 70 km in the southwest to the northeast direction and a width of approximately 20 km. The most intense simulated rainfall area was located on the east side of Hiroshima city and was shifted eastward by only around 20 km compared with JMA radar observations. Figure 5c shows a three-dimensional visualization of the cloud distribution over Hiroshima at 00:10 am viewed from the southeast direction. It was visualized using the open software named VDVGE (Kawahara et al. 2015). For this visualization, we ran a 100 m grid-spacing simulation for 10 min starting at 12 am from a linearly-interpolated 500 m grid-spacing model state. The total mixing ratios of non-sedimenting hydrometeors—i.e., cloud water and cloud ice—are shaded in white, whereas those of sedimenting hydrometeors—i.e., rain, graupel and snow—are colored in red. The images are stretched by a factor of three in the vertical direction to aid recognition of the convective system. This visualization shows a typical three-dimensional shape of a storm training, consisting of a chain of cumulonimbus lines.

3.2 Effect of Turbulence Enhancement on Orographic Precipitation

The turbulence-enhanced collision rate of cloud droplets may explain the rapid growth of cloud droplets, which often result in fast rain initiation in the early stages of cloud development (Falkovich and Pumir 2007; Grabowski and Wang 2013).

The change rate of particle number density, n_f (r, x_i, t), by the stochastic collision-coalescence process is represented by

$$\frac{d{n}_{f}(r)}{dt}=\frac{1}{2}\underset{0}{\overset{\infty }{\int }}{K}_{c}\left({r}^{{\prime}}, {r}^{{{\prime}}{^{\prime}}}\right){n}_{f}\left({r}^{{\prime}}\right){n}_{f}({r}^{{{\prime}}{^{\prime}}})dr{^{\prime}}- \underset{0}{\overset{\infty }{\int }}{K}_{c}\left(r,r{^{\prime}}\right){n}_{f}\left(r\right){n}_{f}\left({r}^{{\prime}}\right)dr{^{\prime}}$$

(1)

where ${r}^{{{\prime}}{^{\prime}}}={\left({r}^{3}-{r}^{{{\prime}}3}\right)}^{1/3}$ and ${K}_{c}\left({r}_{1}, {r}_{2}\right)$ is the collision kernel describing the rate at which a particle of radius r₁ is collected by a particle of radius r₂. The conventional collision kernel model is the hydrodynamic kernel model, which describes the collision due to the settling velocity difference between two particles of different sizes

$$\langle {K}_{c,hydr}\left({r}_{1}, {r}_{2}\right)\rangle =\pi {R}_{12}^{2}\left|{V}_{p,1}-{V}_{p,2}\right|,$$

(2)

where <> denotes an ensemble average, R₁₂ (=r₁ + r₂) is the collision radius and V_p,i is the settling velocity of particles with radius r_i. This hydrodynamic kernel cannot describe the collisions due to turbulence since it contains no flow parameters.

The turbulent collision kernel that involves the turbulence effects can be written in the following form (Wang et al. 1998).

$$\langle {K}_{c,turb}\left({r}_{1}, {r}_{2}\right)\rangle =2\pi {R}_{12}^{2}\langle {w}_{r}(x={R}_{12})\rangle {g}_{12}(x={R}_{12})$$

(3)

Here ${w}_{r}(x={R}_{12})$ is the radial relative velocity at contact separation and describes the turbulence-enhanced relative velocities of two colliding particles. The term, ${g}_{12}(x={R}_{12})$, is the radial distribution function (RDF) at contact separation, the so-called the “accumulation effect,” which measures the effect of particle preferential distributions. Currently, there are a couple of models available for the turbulent kernel based on Eq. (3). Examples are the Ayala-Wang model (Ayala et al. 2008; Wang et al. 2008) and our developed model (Onishi model; Onishi et al. 2015). This study adopted the latter model to investigate the turbulent collisions on cloud development.

Figure 6 shows the computational domain for mesoscale orographic flow over Mt. Hiei, located between Kyoto and Shiga prefectures. The MSSG simulation with the MSSG-Bin cloud microphysics model was performed for the domain whose domain was 40 km × 20 km × 15 km in the x-(streamwise), y-(spanwise), and z-(vertical) directions. The number of numerical grid points used were 400 × 200 × 48. Horizontal computational grids were regular, and the height-based, terrain-following-coordinate system has been chosen for the vertical direction. The minimum vertical spacing was 10 m in the vicinity of the surface, and the maximum one was 500 m at the top of the domain. The moist air, whose relative humidity was 95%, flew into the domain with the streamwise velocity U0 = 15 m/s toward the east. In the MSSG-Bin model, 33 bins (classes) were used for so-called mass-doubling size resolution. The droplet collision growth was calculated using Eq. (1). Using the collision kernel model of Eq. (2) (RUN-NoT) or that of Eq. (3) (RUN-T).

Fig. 6

Bird-eye’s view of the computational domain. A snapshot of three-dimensional cloud distribution (in white) and the surface precipitation (in blue) are also drawn

Figure 7 shows the surface precipitation obtained from RUN-NoT and RUN-T. It shows a large part of surface precipitation in windward slopes. This means that the local orography governs the present precipitation process. Intense precipitation, over 80 mm/h, is observed in RUN-T, indicating that the precipitation is significantly enhanced by turbulent collisions of cloud droplets.

Fig. 7

Surface precipitation for a RUN-NoT and b RUN-T. Large part of surface precipitation was seen in windward slope. It was clearly shown that the precipitation is enhanced by turbulent collisions of cloud droplets

3.3 Bulk Parameterizations of Turbulence-Aware Cloud Growth

In the previous subsection, the spectral-bin cloud microphysics scheme was used to investigate the role of turbulent collisions of cloud droplets. The spectral bin simulation is computationally expensive and cannot be used for operational weather prediction simulations.

The influence of turbulence on droplet growth has been parameterized for bulk cloud microphysics schemes in Seifert and Onishi (2016) and Onishi and Seifert (2016). Their bulk parameterization has been implemented in MSSG.

The auto-conversion term, denoted by P_cnnr in Fig. 2, is parameterized as

$${P}_{cnnr}=\frac{{k}_{cc,0}}{20{x}^{*}}\frac{\left(\mu +2\right)\left(\mu +4\right)}{{\left(\mu +1\right)}^{2}}{Q}_{c}^{2}{\overline{{x }_{c}}}^{2}\left[1+\frac{{\Phi }_{au}\left(\tau \right)}{{\left(1-\tau \right)}^{2}}\right]{\eta }_{au},$$

(4)

where $\overline{{x }_{c}}$ (= Q_c /N_c, where N_c is the number density of cloud droplets) shows the mean cloud droplet mass, ${x}^{*}$ is the separating mass between cloud and rain drops, and ${\eta }_{au}$ shows the enhancement factor by turbulence. The dimensionless ratio $\tau ={Q}_{r} /({Q}_{c}+{Q}_{r})$ acts as an internal timescale and modulates the auto-conversion rate through the universal function ${\Phi }_{au}\left(\tau \right)$. The enhancement ${\eta }_{au}$ is given by

$${\eta }_{au}=1+\epsilon R{e}_{\lambda }^{p}\left[{\alpha }_{cc}\left(\nu \right)\mathit{exp}\left\{-\left[\frac{\overline{{r }_{c}}-{r}_{cc}\left(\nu \right)}{{\sigma }_{cc}\left(\nu \right)}\right]\right\}+{\beta }_{cc}\right],$$

(4)

where $\overline{{r }_{c}}$ ($={\left({x}_{c} /{\rho }_{w}\right)}^{1 /3}, \mathrm{where }{\uprho }_{w}$ is the water density) is the mean radius of cloud droplets, $\epsilon$ is the energy dissipation rate and $R{e}_{\lambda }={u}^{{{\prime}}}{l}_{\lambda } / \nu ,$ where ${u}^{{{\prime}}}$ is the RMS of velocity fluctuations and ${l}_{\lambda }$ the Taylor microscale and $\nu$ the kinematic viscosity) is the Taylor-microscale-based Reynolds number. In the case of ${\eta }_{au}=1$, the turbulence enhancement is neglected. The turbulence characteristic variables $\epsilon$ and $R{e}_{\lambda }$ can be estimated from grid-scale quantities, assuming local isotropic homogeneous turbulence.

The accretion term, denoted by P_accr in Fig. 2, is parameterized as

$${P}_{accr}={k}_{cr,0}{Q}_{c}{Q}_{r}{\Phi }_{ac}\left(\tau \right){\eta }_{ac}$$

(5)

where ${\eta }_{ac}$ shows the enhancement factor by turbulence and modeled as

$${\eta }_{ac}=1+{c}_{r}\epsilon {\left(\frac{{x}^{*}}{\overline{{x }_{r}}}\right)}^\frac{2}{3},$$

(6)

where $\overline{{x }_{r}}$ is the mean mass of rain drops. In case ${\eta }_{ac}=1$, the turbulence effect is neglected.

Seifert and Onishi (2016) proposed the parameterizations for two-moment methods, where the number densities in addition to the mass mixing ratios are calculated. The mean masses of cloud and rain categories, $\overline{{x }_{c}}$ and $\overline{{x }_{r}}$, respectively, are calculated from the mass mixing ratios ${Q}_{\left\{c/r\right\}}$ and number densities ${N}_{\left\{c/r\right\}}$ as

$$\overline{{x }_{\left\{c/r\right\}}}=\frac{{Q}_{\left\{c/r\right\}}}{{N}_{\left\{c/r\right\}}}$$

(7)

The MSSG-Bulk method in the MSSG is based on a one-moment method, where only mass mixing ratios are calculated. The mean masses of cloud and rain categories are estimated by

$$\overline{{x }_{\left\{c/r\right\}}}=\frac{{Q}_{\left\{c/r\right\}}}{{N}_{\left\{c0/r0\right\}}},$$

(9)

where the cloud number density, ${N}_{c0}$, is set to $7.0\times {10}^{7} {\mathrm{m}}^{3}$ (constant) and the rain number density ${N}_{r0}$ is calculated from ${Q}_{r}$ using the formulation proposed in Thompson et al. (2004).

3.4 Sri Lanka Rainfall Simulation with the Turbulence-Aware Bulk Cloud Parameterizations

The MSSG model was applied to Aranayake heavy rainfall event in Sri Lanka in May 2016, which caused devastating landslides resulting in over 150 casualties.

Figure 8a shows the computational nesting domains. Three-layer nesting domains were used. The horizontal resolutions were 8 km, 2 km, and 500 m. The number of horizontal grids was N_λ × N_φ = 128 × 128 for the 8 km-resolution domain, and N_λ × N_φ = 256 × 256 for the 2 km- and 500 m-resolution domains. Irregular grid spacings were used for the vertical direction with the same number of vertical grids, N_z = 55, for all the three nesting domains for the height of 40 km. The simulation was conducted for three days starting from 00:00UTC 14 May 2016.

Fig. 8

a Computational nesting domains for the Aranayake rainfall simulation. b Rainfall accumulation from 06:00UTC on 15 May 2016

Figure 8b shows the rainfall accumulation at Aranayake (80.4546E and 7.1476 N) from 06:00UTC 15 May, 2016. The simulation results for both Case-NoT (turbulence enhancement not considered) and Case-T (turbulence enhancement considered) cases are drawn together with the observation. Though the simulation results did not reproduce timings of heavy rainfalls, they successfully reproduced the total rainfall at the end. Case-T shows a larger amount of rainfall than Case-NoT, which indicates the turbulence enhancement made an impact on the orographic rainfall. Interestingly, turbulence impact fluctuated from time to time. For example, there is little difference between Case-NoT and Case-T for the rainfall around 09:00UTC on 15 May, while the large difference between the two for the rainfall around 03:00UTC on 16 May. This means that the turbulence enhancement of droplet growth can have a large impact on certain weather conditions, and the present bulk parameterization can represent that impact. Future research is expected to clarify the weather condition under which the turbulence enhancement becomes relevant.

4 Realtime Operational Prediction System

4.1 Operational Prediction System

The operational prediction system consists of pre-processing, weather simulation and post-processing (Fig. 9). Pre-processing includes preparing the initial and boundary data for MSSG forecast simulations, including the data assimilation module that produces the analysis data considering the observations. Downloading the observation data is a part of the data assimilation module.

Fig. 9

Flow chart of an operational real-time weather forecasting system

The MSSG weather simulation runs on various computer systems, either notebook PCs or supercomputers. The computer performance limits the domain size, spatial resolution, and forecast leading time. In the current situation (in the year 2022), the ordinary desktop workstation can run 24 h simulations for a 200 km × 200 km domain with 2 km horizontal resolutions within 1 h. Supercomputers can run 1000 times larger or faster simulations.

In post-processing, the results are analyzed, visualized, and transferred to other systems, such as the early warning system. The visualization cost as well as the storage cost (Kolomenskiy et al. 2021) largely depends on the data size. Two-dimensional visualizations do not require much time, but three-dimensional ones usually do. The post-analysis includes machine learning such as super-resolution, which remap low-resolution prediction maps into high-resolution ones. This super-resolution technology will appear in the following subsection.

4.2 Super-Resolution Simulation System for High-Resolution Prediction Maps

Downscaling techniques are used to upconvert the low-spatial resolution models through dynamical and statistical modeling. Various machine-learning models, including artificial neural networks (Cannon 2011) and support vector machines (Ghosh 2010), have been applied for downscaling. Dong et al. (2014) applied the convolutional neural network (CNN) to the super-resolution (SR) and reported that the CNN-based SR outperforms conventional mapping methods. Vandal et al. (2018) applied a CNN-based SR for climate change projections downscaling and reported the advantage of the CNN-based SR compared to the statistical downscaling and traditional SR. Onishi et al. (2019) applied a CNN-based SR to urban micrometeorology and confirmed its robustness. Onishi et al. (2019) further proposed the SR simulation system that utilizes the CNN-based SR technology. The SR simulation system can realize real-time high-fidelity prediction maps on a desktop computer in local offices.

Figure 10 shows the SR simulation system. HR numerical simulations provide better predictions than those obtained using LR simulations, but they are more computationally expensive than LR ones. The SR simulation system consists of an LR simulation and an SR method that maps the resultant LR prediction images to HR ones. This combination provides predictions as good as those obtained using the corresponding HR simulation with a much lower computational cost.

Fig. 10

Super-resolution (SR) simulation system for operational real-time high-resolution prediction. Instead of performing high-resolution (HR) simulations to obtain HR results, low-resolution (LR) simulations are performed. The obtained LR results are converted into HR ones via SR mapping with a deep convolutional neural network (CNN) trained using the dataset obtained from HR simulations

For example, we assume that the HR is 500 m resolution and LR is 2 km resolution for local orographic rainfall prediction. A 500 m horizontal resolution would be required for resolving local slopes, while such an HR weather simulation would require computational time, hindering real-time operations on consumer computers.

Figure 11 shows the recently proposed SR neural network (Yasuda et al. 2022) that can be used for the rainfall prediction map. Several input channels can be used to obtain the output, i.e., the HR rainfall map. For example, the HR topography map would definitely help to improve the accuracy of the SR. The SR neural network is to be trained by a set of HR and LR simulation results. The HR simulations would be performed on supercomputers at a high cost. In the operational stage, however, the LR simulation requires much less computational cost. This SR system is thus promising for real-time prediction services for local communities, which would not have large computational resources.

Fig. 11

Squeeze-and-excitation super-resolution convolutional neural network (SE-SRCNN) for remapping low-resolution rainfall prediction maps into high-resolution ones. The labels above the rectangles of “Conv2D” display kernel size, number of filters, and stride size; for instance, “k9n64s1” means that the kernel size is 9 × 9, the number of filters 64, and the stride size 1 × 1. The image height h and width w are 64 in the present figure. The SE block is kept even when the number of inputs c is 1

For example, let us assume a two nesting layer simulation with a horizontal resolution of 8 and 2 km with 64 × 64 grid points. It takes 1 h for 25 h prediction on Intel Core i7 quadcore on 2.3 GHz, a common consumer CPU. This simulation provides the rainfall prediction map with a 2 km resolution for a 128 km × 128 km area with 24 h leading time. This 2 km resolution map can be super-resolved (upconverted) into a 500 m resolution map that can resolve individual mountain slopes. It takes only milliseconds for the conversion (i.e., inference), and it would not affect the leading time. In the end, the high-fidelity rainfall prediction map with 500 m resolution for 24 h ahead, which can be useful for local early warning systems, can be provided on a consumer PC.

5 Conclusions

We have described a recent development of our next-generation numerical weather prediction model –the Multi-Scale Simulator for the Geoenvironment (MSSG). MSSG is categorized as a global cloud-resolving model that can be used for both global and regional simulations with high resolution without the aid of cumulus parameterization. The examples have shown that MSSG has good prediction skills for the investigation of detailed orographic local rainfall. MSSG can thus be a promising numerical weather prediction model for early warning of landslides.

MSSG has successfully reproduced heavy rains associated with typhoons and a band-shaped precipitation system (Senjo-Koutai in Japanese). The turbulence-aware cloud microphysics model has been implemented in the MSSG. This can bring the extra ability of reliable predictions of orographic rainfall to MSSG.

In order to reproduce local orographic rainfall, the simulation needs enough fine resolutions with O(100 m) to resolve individual local slopes. In the case of an early warning system for a local community, the system should preferably be compact like a desktop PC. The feasible rainfall forecast simulation will have O(1 km) horizontal resolutions, much coarser than O(100 m) resolutions, due to the limited computational speed. The super-resolution simulation system, which relies on recent machine-learning technologies, can help obtain high-resolution rainfall prediction maps that can resolve individual slopes in short operational cycles.

We now have the MSSG and the super-resolution simulation system. The high-resolution rainfall prediction maps can be obtained in real-time, e.g., 24-h rainfall forecast maps can be updated in an hour. The rainfall prediction maps are then transferred to the landslide simulation model and to the early warning system. We hope this integrated system will help construct a safe society free from victims due to landslides.