Applications

The AI/ML algorithms have vast applications in the Earth Science problems. Figure 2 depicts few applications in the areas, including atmosphere/biosphere, seismology, and ocean. Figure 2: An overview of the application of AI/ML algorithms in few areas from Earth Science problems. The precipitation forecasting can include data from short-range, medium-range and extended-range forecasting.

Figure 2: An overview of the application of AI/ML algorithms in few areas from Earth Science problems. The precipitation forecasting can include data from short-range, medium-range and extended-range forecasting.

Statistical downscaling

Downscaling of data is necessary to get a local projection of the information. The present-day models and observations generated from weather stations (or other instruments) are available at coarser resolution and are irregularly spaced, which may often provide misrepresentation (or absence) of precipitation, temperature, or other variables at local levels. Downscaling of Indian Summer Monsoon (ISM) rainfall is a difficult task as it involves a multi-scale spatiotemporal dynamical process with significant variance18 . Further, regional variations of ISM rainfall are often quite large varying from a few millimeters to thousands of millimeters within a few hundred kilometers. The ISM rainfall can be classified into different coherently fluctuating zones, which may be linked to complex multi-scale processes19–21 . The statistical downscaling is a low-cost method to obtain information at the local scale and provide it to stockholders. Artificial Intelligence (AI) and Machine Learning (ML) techniques are being used for statistical downscaling8,22 . Recently, development in the Single Image SuperResolution (SR) using Deep Learning proved to be one of the best methods used for this purpose8– 10 . Another method that showed promising results in the statistical downscaling is the ConvLSTM method documented by Harilal and coauthors23 .

Seismological events

The growing volume of seismological and other geoscience-related datasets acquired from surface and borehole studies, require efficient techniques for analysis and trend recognition to extract valuable signals. AI/ML tools have been applied in different fields in seismology, from event identification to earthquake prediction, with varying degrees of success24–29 The case studies also bring out the need for further research and development to refine the existing techniques and develop new tools that could be utilised in the processing and analyses of large datasets and identification of different geophysical signals. The use of AI/ML techniques in geoscience / seismological could be employed gainfully to analyse other seismological datasets that are routinely acquired by MoES and its affiliated institutions. Identifying seismic phases, accurately, is one of the primary requirements for seismological data analysis towards the determination of earthquake source parameters. ML is planned to be used in identifying different seismic phases in the data. In many earthquake detection algorithms, STA/LTA criteria are being used to detect possible arrival times of P and S waves30 . Later, matched filtering or template matching technique was used for event detection. In this method, waveforms of known events are used as templates to scan through continuous waveforms to detect new events 31 . Recently, ML is utilized to improve earthquake detection and phase picking capabilities25,32 . Fingerprinting and Similarity Thresholding (FAST) is the latest algorithm using ML techniques to identify earthquakes without prior knowledge of seismicity. Being computationally more efficient than template matching FAST would facilitate automated processing of large voluminous datasets. Similarly, the generalized phase detection (GPD) algorithm searches for near identical waveforms from millions 10 of seismograms, which is used to classify windowed data as P, S or noise. GPD can be applied to datasets that were not just encompassed by training sets but can also be applied to difficult cases such as clipped seismograms. Kong et al., (2018) 33 used Neural Networks (NN) for training to pick P-wave onset and to detect P-wave polarity. ML techniques have important applications in the detection of small magnitude local earthquakes in areas which are characterized by sparsity of receivers. ML/AI algorithms may play an important role in the identification of events and in locating earthquakes with recordings of the events at fewer stations33 . Other applications in earth sciences such as hydrology, AI/ML can be used for estimating and predicting stream flow in ungauged basins35–37 .

Short and medium range forecasting using machine learning

Currently, the global highest resolution ensemble predication system at ~12.5 km horizontal resolution (with 21 members) is being used for providing 10 days probabilistic forecast based on the Global Ensemble Forecast System (GEFS@T1534) India Meteorological Department. The high resolution GEFS has been implemented by IITM for operational application since June 2018. While the deterministic GFS model38 at 12.5 km provides a better skill up to ~5 days compared to the earlier coarser resolution (~25 km resolution GFST574)39 , the ensemble prediction system has shown much better skill than the control member (the deterministic GFS model) particularly for predicting extreme rainfall events40-41 . The model forecast inaccuracies mainly arise from initial conditions and improper physical parameterizations. The uncertainties of initial conditions are largely resolved by the perturbed initial conditions in the ensemble prediction system, however, the uncertainty arising from deterministic closures of the physical parameterization still adds much errors due to unrealistic constraints, namely the quasi-equilibrium42 . Under the AI/ML/Dl paradigm, the use of sub-grid scale tendencies generated by the cloud resolving models within each climate model grid as the input of a deep learning model for training to target the heat and moisture trained tendencies hold promise in improving the model fidelity43–45 .

Machine learning for extended range forecasts

AI/ML methods recently find applications in the climate forecast models. Two basic applications show promise for near future applications. The first one is the bias correction and improvement of numerical model forecast. The other one is the methods attempting the sub seasonal low frequency forecast. The bias correction and model post processing applications are of much use to the stakeholders using climate forecast. The climate forecasts from the dynamical models show a lot of bias when forecast is considered over scales lower than the balanced flow, mainly arising due to unknown physics or unresolved dynamics. In situations where enough observation are available over a location, some of the systematic errors arising due to unresolved scale dynamics or physics can be corrected46 . Subseasonal forecasting using machine learning methods are recently under active research12,47–49

Machine learning for seasonal and climate scale forecasting

11 Seasonal forecasting is one of the tough problems in forecasting. As pointed out by Lorenz (1963)50 the weather forecasts are strongly dependent on initial conditions (today’s weather determines tomorrow’s weather) and in contrast climate projections/decadal predictions (an average of weather for a few decades) do not suffer from the initial conditions, however, depends on boundary conditions. When we try to make seasonal forecasts, the distinction is rather blurred and the seasonal forecasts still depend on intial conditions51 . Chattopadhyay et al (2016)51 have shown that models hindcasts initialized with February initial conditions exhibit better prediction skill for Indian Summer Monsoon rainfall (ISMR). Further complexities such as resolving ocean processes also become important at seasonal scale. Hence, extracting predictive information (which changes from event to event) across both space and time scales is very important to make significant progress in seasonal forecasts52 . Therefore, use of AI/ML methods for making improved seasonal forecasts is imperative and research community started using these methods widely in seasonal forecasts53–55 and some researchers even believe that AI/ML methods can outperform conventional prediction systems54-55 . One of the long-standing seasonal prediction problems is prediction of Indian summer monsoon rainfall. H.F. Blandford started seasonal forecasting of Indian summer monsoon using empirical methods in 1886. Since then, numerous attempts were made to predict seasonal mean monsoon over India using both empirical models and dynamical models (Atmosphere and coupled oceanatmosphere models, see review article of Rao et al., (2019) 39 for more details). Empirical models showed very high skills (>0.9) during development stages and during actual operational phase they showed weak skills (<0.5). On the other hand, dynamical models showed moderate skill during hindcast period as well as during operational forecast39 . The major reason for empirical models’ failure to provide high skills during operational phase is that the relationship between predictor and predictands undergo secular changes from the time the model was developed to the phase when it is made operational. To avoid such a situation AI/ML models can be used efficiently to identify new predictors53 . Using autoencoders Saha et al (2021)53 have developed an AI/ML model to predict IMSR with two months lead time and absolute mean error less than 3%. On the other hand, the dynamical models exhibit systematic biases in precipitation (See Rao et al., 201939 ) and basically arise due to parametrization schemes used in these models and therefore underestimate extremes. To avoid such systematic problems AI/ML models will become handy.

Machine learning for improving the physical processes in dynamical models

Dynamical models work on the principle of solving partial differential equations over the area of interest with the necessary initial and boundary conditions. They consist of various components such as atmosphere, ocean, land surface and others. The correct representation of physical processes in the numerical models is highly essential for accurate simulations of the coupled climate system. For example, various researchers have tried to understand the relationship of global and regional teleconnections such as ENSO56-57 , IOD58 , North Atlantic Oscillation59 , Pacific Decadal Oscillation60 , volcanic eruptions61 , aerosols62-63 to Indian monsoon. Recent studies have attempted the use of deep learning to develop models which better represent the physical processes. For example, Witt et al (2019) 64 used deep reinforcement learning based approach to test the stratospheric aerosol injection on climate. Volcanic eruptions have been used 12 as an analog for the stratospheric aerosol injection and deep learning can assist in addressing the non-linear nature of the problem. Recently Lamb et al (2021) 43 used graph neural networks to learn the aerosol optical properties. Similarly, Seifert et al (2020) 65 discuss the role of machine learning in estimating the cloud microphysics. The uncertainties in the simulation of Indian monsoon arise from the missing or erroneous physics in the dynamical systems. Machine learning to improve the physical processes can lead to cascading returns by improving the hydrological outputs from the numerical weather prediction models66–70 .

Machine learning for Nowcasting and tracking the storms cells

The need for a high-resolution early warning system with reliable nowcasts in the regions of steep topography and urban areas during severe weather is highly essential. Traditionally, Nowcasting is served by carrying out extrapolation, probabilistic Nowcasting71 , semi-Lagrangian advection scheme72 , and using algorithms like optical flow etc. The latest data-driven approach is playing a key role in the area of Nowcasting too. Doppler Weather Radar provides extremely high geographical and temporal resolution weather information. Agarwal et al. (2019) 73 utilised radar pictures to forecast the weather using the U-Net algorithm, demonstrating that it outperformed the optical flow technique. Su et al., (2020) 74 have shown that machine learning approaches have a high learning capacity and enhance echo position and intensity forecast accuracy in convective cells. The temporal precision of such convective cells varies from 30 to 60 minutes during a relatively short period of time. The underestimating of precipitation in complicated orography regions is a well-known problem in precipitation estimation. Arulraj and Ana, (2021) 75 used detection and classification machine learning algorithms to improve orographic precipitation across the Southern Appalachian Mountains. Machine Learning for NWP. Satellite remote sensing and NWP groups are ripe for rapid advancement in their application of machine learning. The NWP relies heavily on the integration of fields generated by satellites and other remote sensing devices. Gaps, both spatially and temporally, are a common occurrence in such data. The existence of gaps, both spatially and temporally, is a typical issue in such observations. The time series of satellite ocean fields are constructed using an ensemble of NNs with varying weights76 , and a deep learning method to reconstruct the optical images77 . When modeling, deploying systems, and even issuing warnings, the ML method can give a post-forecast correction to account for uncertainties after learning from all previous failures78 .