How to make an accurate prediction of climate change and greenhouse carbon emissions with artificial intelligence? We are experimenting with several AI models for climate change and global warming. Sri Amit Ray explains the hybrid system of process-based earth system models and the deep learning networks of artificial intelligence.
As climate change and global warming have become the most urgent issues for human survival, researching ways to improve climate change and earth system models has become of the utmost importance.
Artificial intelligence (AI), specifically deep learning algorithms, has the ability to make decisive interpretations of large amounts of complex data. Moreover, modern machine learning techniques appear as the most effective tool for the analysis and understanding of ocean, earth, ice formations, CO2 emissions, biodiversity, and atmospheric data for situation and target specific dynamic prediction. One of the strengths of machine learning tools such as convolutional neural network is its capacity to combine a variety of methods.
Deep Transfer Learning Models for Biodiversity, Climate Change, and Global Warming
Earth system models: An overview
Earth system models (ESMs) are primarily based on general circulation models (GCMs). They are largely used in current climatic studies that take into account features such as biogeochemical cycles and atmospheric chemistry. Earth system models can integrate the influence of human decisions, making them excellent tools for planning infrastructure, energy production and consumption, and landscape use.
The models incorporate the physical, chemical, and biological processes that all work together to shape our planet and the organisms that live on it. Plants respond to temperature changes and rainfall by adjusting the carbon and radiation balance in the atmosphere. These models are based on physical laws such as mass, energy, and momentum conservation. Because they forecast climate using physical boundary conditions, these models are known as “bottom-up approaches.”
The earth’s natural system is a very dynamic and complex system. Traditional ESMs can predict the evolution of the weather on a daily time scale with some accuracy. They are, however, unable to accurately predict climate change over a period of months or years. Seasonal meteorological forecasting, forecasting extreme events such as flooding or fire, and long-term climate projections remain significant challenges.
Currently, the majority of climate forecasting research efforts are based on mechanistic, bottom-up approaches such as physics-based general circulation models. However, the climate is shaped by many biological, human social systems, and chemical processes in addition to the physics of radiation and fluid dynamics.
AI Components of the Earth System Models
In the atmosphere, gases such as water vapour, carbon dioxide, ozone, and methane act like the glass roof of a greenhouse by trapping heat and warming the planet. These gases are called greenhouse gases. Deep transfer learning models are used for different green house effects modeling. The greenhouse effects and the earth system AI models consists of the following ten main AI components:
- Land model
- Atmospheric model
- Ocean model
- Sea ice model
- Land ice model
- Aerosol model
- Carbon cycle model
- Dynamic Vegetation model
- Biogeochemistry model, and the
- Human influences model.
Evolution of the earth system models
The evolution of climate change and the earth system models are shown in the figure below.
Evolution of the climate change and the earth system models.
With the rapid growth of the Internet of Things (IoT), satellite and radar imaging technologies, sensor technologies, and deep learning algorithms of AI, it is possible to develop climate models, which integrates remote sensing (RS), Geographical Information Systems (GIS), and Global Positioning Systems (GPS).
The land components
The land component includes precipitation and evaporation, streams, lakes, rivers, and runoff as well as a terrestrial ecology component to simulate dynamic reservoirs of carbon and other tracers.
Oceanic components
The oceanic component includes features such as free surface to capture wave processes; water fluxes, or flow; currents; sea ice dynamics; iceberg transport of freshwater; and a state-of-the-art representation of ocean mixing as well as marine biogeochemistry and ecology.
The atmospheric component of the ESMs includes physical features such as aerosols (both natural and anthropogenic), cloud physics, and precipitation. GCMs represent layer clouds in the middle and upper troposphere through explicitly-resolved vertical motions and large-scale moistening.
Biogeochemical cycles and terrestrial ecosystems
A biogeochemical cycle is the movement of an element or a compound, such as carbon, nitrogen, or water, between its various living and nonliving forms and locations. Incorporating biogeochemical processes such as the carbon cycle and human influence on atmospheric and ocean warming is vital for climate models. Over geological time, biogeochemical cycles are responsible for altering the chemistry of the ocean, atmosphere, and terrestrial ecosystems.
Aerosols and gases
Aerosol particles are very minute, suspended particles in the air. They have an impact on the Earth’s climate by directly reflecting and absorbing atmospheric radiation and indirectly serving as cloud condensation nuclei and changing the optical properties of clouds.
As both light and heat energy pass through the earth’s atmosphere they encounter the aerosols and gases surrounding the earth. These can either allow the energy to pass through, or they can interrupt it by scattering or absorption. If the atoms in the gas molecules vibrate at the same frequency as the light energy, they will absorb the energy and not allow it to pass through.
Biomass burning
Biomass burning aerosols make up a majority of primary combustion aerosol emissions [1]. The global aerosol model simulates the distribution of aerosols. It includes emissions, atmospheric transport, chemical reactions, gravitational settling, removal processes by dry deposition and precipitations.
Paris Agreement and Artificial Intelligence
The Paris Agreement went into effect on 4 November 2016, which requests all parties to reduce carbon emissions to well below 2 °C and enhance adaptive capacity, and has led to enhance the importance of mitigation and adaptation measures to climate change. Making accurate climate change projections and revealing the mechanism of climate change is the foundation of all climate change measures.
Artificial intelligence for Climate Change
The traditional climate models use a set of well defined differential equations to simulate the atmosphere, the ocean, sea, ice, the land surface, deforestation and vegetation on land, and the biogeochemistry of the ocean. The main limitation to using models is that they are dependent on the accuracy of the underlying model structure. They often have imperfect knowledge, insufficient data, and imperfect equations.
Advance Artificial Intelligence (AI) techniques like deep reinforcement learning, Monte carol adaptive learning, and self-reflection deep learning techniques are very promising tool to solve climate change challenges.
Traditional machine learning techniques like kNN, K-Means, Random Forest, SVM, Convolutional Neural Networks (CNN), Auto-associative Neural Network, Isometric Feature Mapping, Recurrent Neural Networks (RNN), Recursive Neural Networks, Q-learning, Transfer Learning, Long Short-Term Memory (LSTM) are also very useful for certain areas of climate change modeling.
AI for understating the climate change on biological species
Our deep learning models predict that as temperatures rise, the number and diversity of species, which define biodiversity, will decline dramatically. Climate change will also have an impact on plant and animal life as it alters temperature and weather patterns.
Artificial Intelligence for Biodiversity and Climate Change Models
Pollution has a direct impact on animals because it affects the food they eat and the habitats they live in. The reinforcement learning models are developed to model biodiversity and to simulate the adaptability dynamics to new patterns of temperature and rainfall. This module includes:
• managing habitats for endangered species,
• creating refuges and buffer zones, and
• establishing networks of terrestrial, freshwater and marine protected areas.
Kyoto Protocol and Artificial Intelligence
The Kyoto Protocol is a protocol to the United Nations Framework Convention on Climate Change, adopted in Kyoto, Japan, in 1997, commits 37 industrialized countries and the European Union to the so-called Kyoto target of reducing their greenhouse gas emissions by an average of 5% against 1990 levels.
We observed that reinforcement learning is very useful for strategists in running simulations and experimenting with innovation ideas for reducing the greenhouse gas emissions, and evaluating the outcome of different strategic processes.
Generative Adversarial Networks for Climate Change
We study the applicability of Generative Adversarial Networks (GANs) for generating synthetic data related to climate change, biodiversity, and CO2 emissions. We used four main architectures of GAN and trained over 84 GAN models and generated synthetic data with them. We used both 1D time-series data as well as 2D images.
Our climate change GAN models are made up of two sub-systems:
- Generator part generates new data instances from random satellite images of earth.
- Discriminator part is trained on both real data and generated data (from the generator).
- Then, it evaluates whether the input data is real or false.
These two sub-models compete with one another: the discriminator improves its ability to distinguish between generated data and real data, while the generator improves its ability to generate more realistic data points. This process is repeated until the generator is able to generate data instances that the discriminator is unable to distinguish from real data. The input random data includes aerial photographs, satellite imagery, infrared images of the earth surface temperature, and other images and time series data.
Conclusion:
We observed that using “human-in-the-loop” in decision making, strategy evaluation, and benchmark fixing for climate change is crucial. Combining domain expertise, climate physics with AI has also been shown to reduce algorithmic bias. In general, it is important to remember that an algorithm’s ability to make powerful predictions is only as strong as the quality of the data fed to it and how well it fits the processing task.
The algorithms of deforestation, biodiversity, and cloud physics frequently feed off of each other’s actions. The output of one algorithm for climate change frequently influences the learning of another. As we have seen with high-frequency weather model data, this interconnectedness can sometimes lead to erratic behavior in the models, which needs further tuning and adjustments.
For time-series data, we combined a feature selection–based extreme gradient boosting (XGBoost) model with a deep learning–based LSTM model. Making use of various datasets, we are able to build a machine learning model to predict the trends in long-term climate change. Reinforcement learning models are useful for biodiversity strategy analysis.
We observed that deep learning models learn to predict weather patterns as well as climate change patterns directly from observed data rather than incorporating explicit physical laws, and they can compute predictions faster than physics-based techniques.