Morning session
Understanding weather and climate datasets
Explore station observations, satellite products, reanalysis, operational forecasts and climate-model outputs. Compare their strengths, limitations, spatial coverage and temporal resolution.
Day 2 – Climate Data
| Site: | ACMAD eLearning |
| Course: | Summer School on AI for Climate Modelling and Forecasting 2026 |
| Book: | Day 2 – Climate Data |
| Printed by: | Guest user |
| Date: | Saturday, 26 September 2026, 11:22 AM |
Description
1. Day 2 Overview
African Summer School · Day 2
Climate Data: From Observations to Model Outputs
Weather and climate applications depend on many different types of data. Today you will explore where these datasets come from, what they represent, how they are stored and how they can be prepared for analysis.
Day 2 programme
Afternoon session
Data formats, preprocessing and Python
Learn how weather and climate information is stored in CSV, NetCDF and GRIB files. Practise inspecting variables, dimensions, units, missing values and metadata before using the data in an AI workflow.
Learning outcomes
By the end of Day 2, you should be able to:
1. Identify data sources
Distinguish station, satellite, reanalysis, forecast and climate-model datasets.
2. Compare datasets
Explain differences in coverage, resolution, frequency, uncertainty and accessibility.
3. Recognise data formats
Recognise the basic characteristics of CSV, NetCDF and GRIB files.
4. Inspect data quality
Check variables, units, coordinates, missing values, time periods and metadata.
5. Prepare data for analysis
Describe the main preprocessing steps required before training an AI or ML model.
6. Explore data with Python
Open a dataset, examine its structure and create a simple visualisation.
Today’s data journey
Every successful weather or climate AI application begins with understanding the data before selecting a model.
1. Discover Find relevant data sources
2. Understand Read variables and metadata
3. Check Assess quality and gaps
4. Prepare Clean and transform
5. Explore Analyse and visualise
Opening reflection
Think about the data you currently use in your organisation or country.
Consider:
- Which weather or climate datasets do you use most frequently?
- Where do these datasets come from?
- What limitations or data gaps affect your work?
- Which additional datasets would improve your services?
2. Inauguration and Keynote
Day 1 · Morning Session
Inauguration and Keynote
Explore why artificial intelligence is becoming increasingly important for African weather forecasting, climate services, disaster-risk reduction and early warning.
Purpose of the opening session
The opening session introduces the Summer School objectives, facilitators and participants. It establishes the connection between artificial intelligence, operational meteorology, climate modelling, disaster-risk reduction and impact-based early warning.
Participants are encouraged to consider both the opportunities offered by AI and the scientific, technical and institutional conditions required for its effective and responsible use.
Keynote themes
Growing data volumes Satellite, station, reanalysis and forecast datasets are increasing in volume and complexity.
Local forecast needs Users require timely and locally relevant information for preparedness and decision-making.
Data-sparse regions AI may help combine multiple data sources where observations are incomplete or unevenly distributed.
Human expertise AI should complement physical understanding and professional forecasting judgement rather than replace them.
Watch
Opening keynote video
While watching, identify one opportunity for AI in your institution and one challenge that would need to be addressed before operational use.
A transcript or summary should also be provided for participants who cannot access the video.
While watching
- Identify one weather or climate service that could benefit from AI.
- Identify the intended users of that service.
- Note the datasets that might be required.
- Record one technical or institutional barrier.
- Consider the role that forecasters should continue to play.
Explore
Why is AI relevant to weather and climate services?
Processing complex datasets
AI methods can analyse large collections of station observations, satellite imagery, model forecasts and climate records more rapidly than would be possible through manual analysis alone.
Identifying patterns
Machine-learning models can identify statistical relationships in historical data that may be useful for prediction, classification or forecast correction.
Improving local information
AI may help produce more locally relevant information from coarse-resolution forecast and climate-model products.
Supporting operational decisions
AI can help organise, prioritise and present information, but qualified professionals should continue to interpret the evidence, communicate uncertainty and remain accountable for warnings.
Small-group discussion
Select one weather or climate challenge faced by your institution and discuss:
- Why the challenge is important.
- Who is affected by it.
- What information is currently available.
- Where AI might contribute.
- What limitations or risks should be considered.
Keynote presentation
Review the presentation slides and use them to support the group discussion.
View the Keynote Slides3. Understanding AI, Machine Learning and Deep Learning
Day 1 · AI/ML Fundamentals
Understanding AI, Machine Learning and Deep Learning
Learn how artificial intelligence, machine learning and deep learning are related, how they differ, and how they can support weather and climate applications.
Learning objectives
By the end of this chapter, you should be able to:
- Explain artificial intelligence in simple terms.
- Describe machine learning as a branch of artificial intelligence.
- Explain how deep learning relates to machine learning.
- Differentiate training, validation, testing and inference.
- Recognise examples of AI methods used in weather and climate services.
How the three concepts are related
Artificial intelligence is the broadest concept. Machine learning is one approach within artificial intelligence. Deep learning is a family of machine-learning methods based on neural networks with multiple processing layers.
Artificial Intelligence The broad field of systems that perform tasks involving prediction, pattern recognition, reasoning or decision support.
Machine Learning A branch of AI in which models learn patterns from examples in data.
Deep Learning A type of machine learning based on multilayer neural networks.
Artificial intelligence
Artificial intelligence refers to computer systems designed to perform tasks that normally require human intelligence. These tasks may include recognising patterns, interpreting images, processing language, generating predictions and supporting decisions.
AI does not refer to one single method. It includes rule-based systems, optimisation, expert systems, machine learning, neural networks and other computational approaches.
Weather example A computer system analyses satellite images and identifies rapidly developing convective clouds that may produce severe weather.
Machine learning
Machine learning is a branch of artificial intelligence in which a model learns relationships from historical examples instead of relying only on rules written manually.
During training, the model examines input data and known outcomes. It estimates patterns that can later be used to make predictions for new cases.
Weather example A model learns from past numerical forecasts and station observations to correct systematic temperature forecast errors.
Deep learning
Deep learning is a type of machine learning based on artificial neural networks with multiple processing layers. These layers allow the model to learn increasingly complex representations of the data.
Deep-learning approaches are often used for large, complex datasets such as satellite imagery, radar sequences and gridded weather fields.
Weather example A neural network analyses a sequence of satellite images to estimate the movement and development of clouds and rainfall.
Common machine-learning approaches
Supervised learning The model learns from inputs paired with known answers, such as historical forecasts paired with observed temperatures.
Unsupervised learning The model searches for patterns or groups in data without being given predefined answers.
Reinforcement learning A system learns through repeated actions and feedback based on rewards or penalties.
Training, validation, testing and inference
Training The model learns relationships from historical examples.
Validation A separate dataset helps compare models and adjust settings.
Testing Independent data are used to estimate final model performance.
Inference The trained model is applied to new data to produce a prediction.
Avoid data leakage
Information from the test period must not be used during model training. Otherwise, the reported performance may appear better than the model’s true ability to work on unseen data.
Explore
Which concept best fits each example?
A system uses manually defined rules to issue a warning when rainfall exceeds a fixed threshold.
This is an example of a rule-based AI system. It does not necessarily use machine learning.
A model learns from historical forecasts and observations to predict forecast error.
This is machine learning, most likely supervised learning.
A multilayer neural network processes satellite images to predict rainfall.
This is deep learning because it uses a multilayer neural network.
A trained model receives today’s forecast data and produces tomorrow’s temperature prediction.
This is inference: the trained model is being applied to new data.
Quick self-check
Is every AI system based on machine learning?
No. AI also includes rule-based systems, expert systems, optimisation methods and other approaches.
Is every machine-learning model a deep-learning model?
No. Machine learning also includes linear regression, decision trees, random forests, gradient boosting and many other methods.
Why should test data be kept separate?
Separate test data provide a more realistic estimate of how the model will perform on examples it has not seen previously.
Does a more complex model always perform better?
No. A complex model may overfit, require more data and computing resources, and be harder to explain and maintain.
4. The Machine-Learning Workflow
Day 1 · From Problem to Prediction
The Machine-Learning Workflow
Follow the main stages used to design, train, evaluate and operate a machine-learning system for weather and climate applications.
Learning objectives
By the end of this chapter, you should be able to:
- Describe the main stages of a machine-learning workflow.
- Connect an operational problem with appropriate data and model outputs.
- Explain why data preparation and independent evaluation are essential.
- Recognise common workflow risks, including data leakage and overfitting.
- Explain why operational monitoring must continue after deployment.
Start with the problem, not the model
A successful machine-learning project begins with a clearly defined weather or climate problem, the people who will use the information, and the decision that the system is expected to support. The model should be selected only after these requirements are understood.
The six-stage workflow
1
Define the problem Identify the prediction target, intended users, required lead time, geographical scale and operational decision.2
Collect the data Obtain relevant observations, satellite products, reanalysis, forecast data, climate-model output and supporting information.3
Prepare the data Check quality, handle missing values, align dates and locations, standardise units and create model-ready inputs.4
Train the model Use historical examples to estimate relationships between the inputs and the target variable.5
Evaluate the model Compare predictions with observations using appropriate metrics, independent data and relevant baseline methods.6
Deploy and monitor Integrate the model into an operational workflow and continue checking its reliability, data inputs and performance.Explore each workflow stage
1. Define the operational problem
Clearly state what should be predicted and how the prediction will be used. A vague goal such as “use AI to improve forecasting” is not sufficient.
A stronger definition would be: “predict whether daily maximum temperature will exceed 40°C at a specific location five days in advance to support heat-preparedness decisions.”
2. Identify and collect the required data
The required data depend on the problem. Possible sources include station observations, satellite products, reanalysis, numerical forecasts, climate-model output, elevation and calendar information.
Data should cover a sufficiently long period and represent the locations, seasons and extreme events where the model will be used.
3. Prepare and explore the data
Data preparation may include checking missing values, removing or flagging invalid observations, converting units, matching forecast lead times and aligning grids with station locations.
Basic exploration should examine distributions, seasonal patterns, extreme values, relationships between variables and geographical differences.
4. Select and train the model
Select a model that is appropriate for the prediction task, data volume, available computing resources and required interpretability.
Begin with a simple baseline before testing more complex methods. Model complexity should be justified by meaningful improvements in performance or operational value.
5. Evaluate performance
Evaluation should use data that were not used for model training. The model should be compared with observations and with relevant baseline forecasts.
Evaluation should consider average performance, extreme events, geographical differences, uncertainty and whether improvements are useful to operational users.
6. Deploy, communicate and monitor
Deployment involves more than saving a trained model. The system needs reliable data feeds, automated processing, quality checks, documentation and clear responsibilities.
Performance should be monitored because data sources, instruments, forecasting systems, user needs and climate conditions can change.
Worked example: improving a temperature forecast
- Define: Improve daily maximum-temperature forecasts for a local station at a five-day lead time.
- Collect: Obtain historical model forecasts, station observations, lead-time information and seasonal variables.
- Prepare: Match forecasts and observations by date, location and valid time; check units and missing values.
- Train: Fit a model using historical forecast inputs and observed maximum temperature.
- Evaluate: Compare corrected and raw forecasts on dates not used during training.
- Operate: Generate daily corrected forecasts and monitor performance, particularly during extreme heat.
Common workflow risks
Poor problem definition The model output may not support a real user decision.
Data leakage Information from the test period accidentally enters model training.
Overfitting The model memorises training data but performs poorly on new cases.
Unrepresentative data Important locations, seasons or extreme events are missing.
Inappropriate metrics Evaluation does not reflect the intended operational use.
No monitoring Changes in inputs or performance remain undetected after deployment.
Check your understanding
Which workflow stage is involved?
Converting rainfall measurements from millimetres to consistent units
This belongs to data preparation.
Deciding that the system must predict dangerous heat five days ahead
This belongs to problem definition.
Comparing predictions with station observations from an unseen year
This belongs to model evaluation.
Checking whether a daily satellite-data feed has stopped updating
This belongs to deployment and operational monitoring.
Institutional reflection
Think about a possible AI application in your institution. Which workflow stage would be the most difficult to complete successfully?
Consider data availability, technical skills, computing infrastructure, evaluation, operational integration and long-term maintenance.
5. AI Applications in Weather and Climate
Day 1 · From Concepts to Applications
AI Applications in Weather and Climate
Explore how artificial intelligence and machine learning can complement observations, physical models and expert judgement across weather forecasting, climate services and early warning.
Learning objectives
By the end of this chapter, you should be able to:
- Identify major AI applications in weather and climate services.
- Connect different applications with appropriate data sources.
- Explain how AI can complement numerical models and observations.
- Recognise where human expertise remains essential.
- Select a possible AI application for an institutional challenge.
AI complements existing forecasting systems
AI should not be viewed as a replacement for observations, numerical weather-prediction models, climate models or professional expertise. Its strongest role is often to combine information, identify patterns, correct systematic errors and produce information that is more useful for operational decisions.
Major application areas
Forecast post-processing Use historical forecasts and observations to improve raw numerical weather-prediction outputs.
Bias correction Identify and reduce persistent differences between model forecasts and observed conditions.
Statistical downscaling Produce more local information from weather or climate-model outputs available at coarse spatial resolution.
Nowcasting Use recent satellite, radar and observational data to estimate weather development over the next few minutes or hours.
Extreme-event monitoring Detect conditions associated with extreme heat, heavy rainfall, drought, severe storms or strong winds.
Data quality control Identify missing, inconsistent, duplicated or potentially erroneous observations.
Climate prediction Learn relationships between large-scale climate drivers and regional rainfall, temperature or drought conditions.
Impact-based early warning Combine hazard forecasts with exposure, vulnerability and potential impacts to support preparedness.
Explore the applications
Forecast post-processing and bias correction
Numerical weather-prediction models can show systematic errors at particular locations, seasons or forecast lead times. A machine-learning model can learn these historical relationships and adjust new forecasts.
Example: A forecast model consistently predicts daily maximum temperature two degrees too low during the hot season. A correction model can estimate and reduce this cold bias.
Statistical downscaling
Global forecasts and climate projections are often too coarse to represent local terrain, coastlines, cities or station conditions. Statistical downscaling learns relationships between large-scale conditions and local observations.
Example: A regional climate model provides rainfall information on a large grid, while district planners require more local estimates.
Nowcasting
Nowcasting focuses on the immediate future, often from a few minutes to several hours. Machine-learning systems can analyse sequences of satellite or radar images to estimate movement and development.
Example: Satellite imagery is used to predict where rapidly developing convective clouds may move during the next two hours.
Extreme-event monitoring
AI can help identify combinations of variables associated with dangerous conditions. However, rare extremes may be poorly represented in the training data and require careful evaluation.
Example: A model estimates the probability that daily maximum temperature will exceed a dangerous threshold five days ahead.
Impact-based early warning
Hazard information alone does not describe the likely consequences. Impact-based warning combines the forecast hazard with information about exposed people, livelihoods, infrastructure and vulnerability.
Example: A heat forecast is combined with settlement location, population vulnerability and access to cooling or health services.
Connecting applications with data
| Application | Possible input data | Possible output |
|---|---|---|
| Bias correction | Historical forecasts and observations | Corrected temperature or rainfall forecast |
| Nowcasting | Satellite images, radar and recent observations | Short-term storm or rainfall prediction |
| Downscaling | Coarse model output, local observations and elevation | Higher-resolution local prediction |
| Extreme-heat monitoring | Temperature forecasts, observations and vulnerability data | Probability or warning category |
| Data quality control | Station records and neighbouring observations | Flags for suspicious or missing measurements |
Check your understanding
Match the challenge with an AI application
A global weather model consistently predicts temperatures that are too low at a local station.
Appropriate application: forecast post-processing or bias correction.
Satellite images show rapidly developing clouds, and guidance is needed for the next hour.
Appropriate application: satellite-based nowcasting.
Climate projections are available on a coarse grid, but planners need district-level information.
Appropriate application: statistical downscaling.
Health authorities need to know which communities face the greatest danger during an extreme-heat event.
Appropriate application: extreme-event monitoring combined with impact-based early warning.
The continuing role of forecasters
Forecasters bring physical understanding, local knowledge and operational experience that may not be represented fully in the training data.
- Review AI outputs alongside observations and physical models.
- Identify unusual situations outside the model’s experience.
- Interpret uncertainty and conflicting guidance.
- Communicate warnings in language appropriate to users.
- Remain accountable for operational decisions.
Group activity: select an institutional application
Select one weather or climate challenge faced by your institution. Discuss the following questions:
- What operational problem needs to be addressed?
- Who would use the resulting information?
- Which AI application is most appropriate?
- Which datasets would be required?
- What benefit could the system provide?
- What limitation or risk would need to be managed?
- What role would human experts continue to play?
6. Opportunities and Limitations in the African Context
Day 1 · African Context
Opportunities and Limitations in the African Context
Examine where AI can strengthen African weather and climate services, while recognising the data, infrastructure, capacity and institutional limitations that must be addressed.
Learning objectives
By the end of this chapter, you should be able to:
- Identify opportunities for AI in African weather and climate services.
- Explain how data scarcity can affect model development and performance.
- Recognise infrastructure, skills and sustainability challenges.
- Explain why model performance must be assessed across locations and user groups.
- Identify practical measures that can reduce implementation risks.
Technology must respond to local realities
AI can provide important benefits, but a model developed with data, infrastructure and institutional conditions from one region may not work equally well elsewhere. Successful applications should be based on local needs, local evaluation and sustained institutional ownership.
Opportunities
Combining multiple data sources AI can combine station observations, satellite products, reanalysis and numerical forecasts to provide more complete information.
Supporting data-sparse regions Satellite and reanalysis products may provide useful spatial coverage where station networks are limited.
Improving local forecasts Bias correction and downscaling can help produce information that is more relevant to specific communities and institutions.
Faster analysis Automated systems can process large volumes of observations, model outputs and satellite imagery more quickly.
Strengthening early warning AI may support earlier detection of high-impact heat, rainfall, drought and severe-weather conditions.
Regional collaboration Shared tools, datasets and training can strengthen cooperation among national services, regional centres and universities.
Limitations and risks
Sparse observations Short, incomplete or uneven station records can limit what a model learns and make evaluation difficult.
Data quality problems Missing values, inconsistent units, station movement and instrument changes can introduce errors.
Geographical bias Models may perform best in areas with dense observations and poorly in remote or under-observed regions.
Limited generalisation A model trained in one country or climate zone may not transfer successfully to another.
Infrastructure constraints Reliable electricity, computing, internet connectivity and operational data feeds may not always be available.
Skills and maintenance Institutions need staff who can evaluate, operate, troubleshoot and update the system over time.
External dependence Reliance on unavailable source code, foreign platforms or short-term projects can threaten long-term sustainability.
Uncertainty and false confidence A precise-looking prediction may be interpreted as certain even when the model has limited evidence.
How data scarcity affects AI models
Important weather situations may be missing
A short training record may contain very few examples of extreme heat, heavy rainfall or severe storms. The model may therefore perform poorly when these rare but important events occur.
Some locations may be under-represented
Models trained mainly with observations from cities or accessible locations may perform poorly in rural, mountainous, desert or coastal environments.
Evaluation becomes less reliable
When independent observations are scarce, it becomes difficult to determine whether apparent improvements are real and whether the model performs consistently across regions.
Gridded datasets may not represent local conditions fully
Reanalysis, satellite and model datasets provide valuable coverage, but they may contain biases and may not reproduce local extremes or station conditions accurately.
Practical ways to reduce the risks
- Combine station, satellite, reanalysis and forecast datasets carefully.
- Document data quality, missing values and known biases.
- Evaluate performance separately by location, season and event severity.
- Compare AI models with simple and operational baselines.
- Report uncertainty, sample size and limitations clearly.
- Test the system with local forecasters and intended users.
- Prefer systems that local institutions can operate and maintain.
- Continue collecting observations and monitoring performance.
Case study: good national performance, poor local performance
A machine-learning rainfall model shows good average performance across a country. However, it frequently misses heavy rainfall in a region where only a few weather stations are available.
What should the project team investigate?
- Whether the region is sufficiently represented in the training data.
- Whether heavy-rainfall events are present in sufficient numbers.
- Whether satellite or model inputs contain regional biases.
- Whether performance metrics were calculated separately for the region.
- Whether uncertainty is larger where observations are sparse.
- Whether additional observations or local recalibration are required.
- What consequences missed events may have for local users.
Check your understanding
Opportunity or limitation?
Satellite data provide coverage where few weather stations exist.
This is an opportunity, although the satellite data should still be validated and their limitations documented.
A model is trained mainly using observations from major cities.
This is a limitation because rural and remote environments may not be represented adequately.
Regional centres and national services share datasets and software.
This is an opportunity for collaboration, capacity development and more efficient use of resources.
The system cannot operate without continuous external technical support.
This is a sustainability and institutional-ownership limitation.
Institutional reflection
What is the greatest opportunity for AI in your institution, and what is the most important limitation that could prevent successful use?
Consider observations, access to datasets, computing, staff capacity, institutional ownership, operational integration and long-term maintenance.
7. Responsible AI
Day 1 · Responsible and Inclusive Use
Responsible AI
Explore how AI systems can be designed and used in ways that are scientifically credible, transparent, inclusive, secure, accountable and relevant to African weather and climate services.
Learning objectives
By the end of this chapter, you should be able to:
- Explain why responsible AI is important in weather and climate services.
- Identify principles of transparency, fairness, inclusion and accountability.
- Recognise risks related to data ownership, privacy and external dependence.
- Explain why human oversight remains essential.
- Use a responsible AI checklist to assess a proposed application.
Responsible AI begins before model development
Responsible AI is not only about checking a finished model. It begins with deciding which problem should be addressed, who should participate, what data may be used, how performance will be evaluated and who will remain accountable for decisions.
Principles of responsible AI
Local relevance The system should address a clearly defined operational or community need rather than introducing technology without a practical purpose.
Transparency Data sources, assumptions, methods, limitations and intended uses should be documented and communicated clearly.
Fairness Performance should be examined across locations, communities and user groups so that benefits and risks are not distributed unfairly.
Inclusion Forecasters, data providers, decision-makers and affected communities should participate in defining needs and evaluating outputs.
Data stewardship Data ownership, access rights, privacy, quality and appropriate reuse should be respected.
Accountability Clear responsibilities should be established for model approval, operational decisions, warning communication and corrective action.
Reliability and safety The system should be tested under normal and extreme conditions, with procedures for identifying failures and preventing harmful use.
Capacity and sustainability Local institutions should be able to understand, operate, maintain and adapt the system over time.
Why human oversight remains essential
AI models learn from historical data and may fail during unusual events, changing conditions or situations that were poorly represented during training. Human experts provide physical understanding, local knowledge and awareness of operational consequences.
- Review AI output alongside observations and numerical forecasts.
- Identify physically unrealistic or inconsistent predictions.
- Consider local conditions not represented in the model.
- Interpret uncertainty and conflicting guidance.
- Decide whether and how warnings should be issued.
- Remain accountable for operational communication and decisions.
Explainability and communication
Users do not always need to understand every mathematical detail of a model, but they should understand what information was used, what the output means, how reliable it is and when it should not be trusted.
Forecast products should communicate uncertainty, known limitations and the role of expert review rather than presenting predictions as guaranteed outcomes.
Data ownership and stewardship
Who owns and controls the data?
National meteorological services, communities and partner organisations may have different ownership and access conditions. These conditions should be agreed before data are shared or reused.
Is personal or sensitive information involved?
Impact-based services may use population, health, mobility or vulnerability information. Appropriate safeguards and access controls may be required.
Is the data suitable for the intended purpose?
Data collected for one purpose may not be appropriate for another. Quality, coverage, limitations and permitted uses should be assessed.
Will local institutions retain access?
Projects should avoid arrangements in which local institutions provide data but cannot access, evaluate or benefit from the resulting models and products.
Risk of external dependence
A system may appear successful during a funded project but become unusable when external technical support, cloud computing or software access ends.
Sustainability planning should include documentation, source-code access, training, maintenance responsibilities, computing costs, data continuity and procedures for updating the model.
Responsible AI checklist
- Is the operational problem clearly defined?
- Have intended users participated in defining the system?
- Are the data sources, quality and ownership documented?
- Has performance been tested across different locations and conditions?
- Are uncertainty and limitations communicated clearly?
- Can affected users question or challenge the system’s outputs?
- Can the institution operate and maintain the system locally?
- Is there a procedure for detecting and correcting failures?
- Is a qualified person responsible for the final decision?
Scenario: an externally developed flood-warning system
An external organisation develops an AI flood-warning system. The national meteorological service cannot inspect the training data, modify the model or operate it without the external provider.
What concerns should be discussed before operational use?
- Transparency of the model and training data.
- Independent validation under local conditions.
- Ownership and access to data and model outputs.
- Long-term dependence on the external provider.
- Maintenance, continuity and operating costs.
- Responsibility when an alert is incorrect or missed.
- Whether forecasters can review and override the output.
Check your understanding
Which responsible AI principle is involved?
Model performance is reported separately for well-observed and data-sparse regions.
This supports fairness, transparency and geographical evaluation.
Forecasters can reject an output that conflicts with observations and physical understanding.
This is human oversight and accountability.
Communities participate in deciding how heat-warning categories should be communicated.
This supports inclusion, local relevance and participatory design.
The project documents who owns the observations and who may reuse them.
This is data stewardship and governance.
Institutional reflection
Consider one AI application that could be introduced in your institution. Which responsible AI principle would require the most attention?
Think about local relevance, data ownership, transparency, fairness, inclusion, capacity, sustainability and human accountability.
8. Activities and Reflection
Day 1 · Apply and Reflect
Day 1 Activities and Reflection
Consolidate your understanding of AI foundations, complete the Day 1 interactive activities and connect the concepts to your institutional context.
Day 1 summary
Artificial intelligence is a broad field that includes systems capable of prediction, pattern recognition and decision support. Machine learning is one branch of AI in which models learn relationships from data, while deep learning uses multilayer neural networks for complex tasks such as analysing satellite images and gridded weather fields.
AI can support forecast post-processing, bias correction, downscaling, nowcasting, data quality control, extreme-event monitoring and impact-based early warning.
However, operational value depends on suitable data, independent evaluation, local relevance, transparency, institutional capacity, sustainable infrastructure and continued human oversight.
Five key takeaways
Start with the problem Define the operational need, users and decision before choosing a model.
Data determine possibilities Model performance depends on data quality, coverage and representativeness.
Evaluation must be independent Models should be tested on unseen data and compared with appropriate baselines.
AI should complement experts Forecasters remain essential for interpretation, communication and accountability.
Responsibility is essential Transparency, fairness, inclusion, sustainability and data stewardship must be considered.
Complete the Day 1 activities
Use the activities below to review the concepts, test your understanding, exchange experiences with other participants and identify a possible AI application for your institution.
Review
Day 1 presentation slides Review the main definitions, diagrams, examples and responsible AI principles introduced during the sessions.Interact
AI concepts and applications Match AI concepts, weather and climate challenges, datasets and possible applications.Assess
Day 1 knowledge check Test your understanding of AI, machine learning, model workflows, limitations and responsible use.Discuss
Responsible AI in the African context Share one possible AI application and one important limitation or risk relevant to your institution.Apply
Institutional AI reflection Describe an operational problem, possible AI application, required data, expected benefit and one limitation.Group case study: should AI be used?
A national meteorological service wants to develop a system that predicts high-impact rainfall at district level using satellite products, numerical-weather-prediction outputs and available station observations.
Discuss the following:
- What operational problem would the system address?
- Who would use the prediction and what decision would it support?
- Which input data would be required?
- How could missing station observations affect performance?
- How should the model be evaluated before operational use?
- How should uncertainty be communicated?
- What role should forecasters continue to play?
- Who would operate and maintain the system?
Personal reflection
Before completing Day 1, write brief responses to the following questions:
- What new concept was most useful to you today?
- What AI application is most relevant to your institution?
- Which dataset would be essential for that application?
- What is the greatest limitation or risk?
- Which responsible AI principle deserves the most attention?
- What question would you like to explore during the remaining days?
Final self-check
Can you explain these ideas?
What is the difference between AI and machine learning?
AI is the broader field of computer systems that perform tasks such as prediction, recognition or decision support. Machine learning is one approach within AI in which models learn relationships from data.
Why should a simple baseline be evaluated?
A baseline shows whether the machine-learning model provides a real improvement over a simpler or existing operational method.
Why might a model perform poorly in data-sparse areas?
Those locations may be under-represented during training and evaluation, and available gridded products may not reproduce local conditions accurately.
Why is human oversight necessary?
Human experts provide physical understanding, local knowledge, interpretation of uncertainty and accountability for operational decisions and warnings.
Day 1 completion checklist
Before finishing Day 1, confirm that you have:
- Reviewed all eight chapters of the Day 1 Book.
- Watched or reviewed the keynote material.
- Completed the interactive AI activity.
- Taken the Day 1 knowledge-check quiz.
- Contributed to the responsible AI discussion.
- Submitted your institutional reflection.
- Identified one question to carry forward into Day 2.