The current Foot-and-Mouth Disease (FMD) crisis in South Africa has evolved into a national biosecurity emergency. As of early 2026, the virus has been detected in eight of our nine provinces, leading to unprecedented measures, including the cancellation of livestock exhibitions at major events like Nampo 2026. While the Agricultural Research Council (ARC) recently celebrated a breakthrough by resuming local vaccine production for the first time in 21 years, the path to eradication remains steep.

For South Africans, this isn’t just a “farming problem.” It’s a food security and economic shock. Outbreaks have driven up the cost of beef and dairy, threatened the R80 billion livestock sector, and led to severe export bans that drain foreign revenue. However, where there is complexity, there is opportunity for innovation. For Southern African students, the FMD outbreak offers a rich, real-world laboratory for Applied Artificial Intelligence.

20 AI Project Ideas for Southern African Students

Here are 20 project templates designed to help students bridge the gap between computer science and veterinary epidemiology.

1. Farm-Level Early Warning from Symptoms + Movement Risk

• Topic: Predictive Syndromic Surveillance.
• Rationale: Early detection is the only way to “get ahead” of the virus before it jumps between districts.
• Objective: Predict which specific farms are at the highest risk of being the next confirmed case.
• Data: Vet reports, movement permits, local clinic logs, and historical outbreak patterns.
• Methodology: Ingest $\rightarrow$ Label confirmed cases $\rightarrow$ Feature engineering (proximity to outbreaks, recent stock intake) $\rightarrow$ Risk scoring.
• Models: XGBoost or LightGBM for tabular risk classification.
• Ethics: Prevent “digital redlining”, ensure risk scores don’t lead to the unfair exclusion of smallholders from markets without clinical proof.

2. Animal Movement Network Analytics: “Super-Spreader” Detection

• Topic: Graph AI for Bio-containment.
• Rationale: Certain auctions or transport hubs act as “hubs” that amplify a local outbreak into a national crisis.
• Objective: Identify and rank the most dangerous nodes in the livestock supply chain.
• Data: Digital movement permits, auction records, and transport route GPS data.
• Methodology: Build a dynamic directed graph $\rightarrow$ Compute centrality metrics $\rightarrow$Simulate node removal (quarantine).
• Models: GraphSAGE (Graph Convolutional Networks) or PageRank-based centrality.
• Ethics: Transparency in how “high-risk” nodes are identified to prevent commercial bias.

3. Forecasting Outbreak Trajectory Under Specific Interventions

• Topic: Causal Time-Series Forecasting.
• Rationale: Decision-makers need to know if a movement ban will actually work in 4 weeks or if it’s too late.
• Objective: Forecast case counts 8 weeks ahead under “Ban” vs. “No Ban” scenarios.
• Data: Weekly case counts per district, dates of policy implementation, and vaccination rollout volumes.
• Methodology: Baseline forecasting $\rightarrow$ Policy feature injection $\rightarrow$ Scenario backtesting.
• Models: Temporal Fusion Transformers (TFT) or Bayesian Structural Time Series.
• Ethics: Avoid “False Certainty”; models must include clear confidence intervals to prevent misallocation of government resources.

4. Optimal Vaccine Allocation Under Supply Constraints

• Topic: Operations Research + Demand ML.
• Rationale: With local production only starting at 20,000 doses/week for a herd of millions, every dose counts.
• Objective: Determine the “best” geographic allocation of vaccines to minimize total R-value (spread).
• Data: Herd sizes, cold-chain logistics (fridge locations), and current outbreak proximity.
• Methodology: Predict demand $\rightarrow$ Formulate as a Constrained Optimization Problem $\rightarrow$Solve.
• Models: Mixed-Integer Linear Programming (MILP) or Reinforcement Learning (RL).
• Ethics: Equity, ensure communal farmers in remote areas are prioritized alongside large commercial feedlots.

5. Computer Vision: Automatic Lesion Detection from Images

• Topic: CV Diagnostic Support.
• Rationale: Vets are outnumbered. A mobile tool could help a farmer know if they need to call an inspector immediately.
• Objective: Identify FMD-specific lesions (mouth/hoof sores) from mobile phone photos.
• Data: Labeled image dataset of healthy vs. infected livestock (vetted by experts).
• Methodology: Preprocessing $\rightarrow$ Data Augmentation $\rightarrow$ Transfer Learning $\rightarrow$ Mobile Deployment.
• Models: MobileNetV3 or YOLOv8 for real-time lesion localization.
• Ethics: Misdiagnosis risk; the tool must clearly state it is for “Triage,” not “Final Diagnosis.”

6. NLP: Automated Intelligence from Veterinary Notices

• Topic: Text Mining for Situational Awareness.
• Rationale: Vital info is buried in PDFs and social media posts from various provinces.
• Objective: Build a real-time, searchable “Outbreak Map” by scraping text.
• Data: Government bulletins, news articles, and agricultural newsletters.
• Methodology: Web crawling $\rightarrow$ Named Entity Recognition (NER) $\rightarrow$ Event extraction.
• Models: BERT-based NER or GPT-4o for structured data extraction.
• Ethics: Fact-checking; the system must prioritize official sources to avoid amplifying “fake news” rumors.

7. Fraud Detection in Livestock Movement Permits

• Topic: Compliance Analytics.
• Rationale: “Paper” permits are often forged to bypass Disease Management Areas (DMAs).
• Objective: Identify impossible travel times or duplicate permit IDs.
• Data: Digital permit logs and known road-network travel times.
• Methodology: Rule-based baseline $\rightarrow$ Anomaly detection for outliers.
• Models: Isolation Forest or Autoencoders.
• Ethics: Due process; ensure flagged “anomalies” are investigated by humans before penalties are applied.

8. Remote Sensing: Detecting High-Risk Grazing Congregation

• Topic: Geospatial AI.
• Rationale: Communal grazing near water points facilitates virus transmission between herds.
• Objective: Map “hotspots” of animal congregation using satellite imagery.
• Data: Sentinel-2 satellite imagery, NDVI (vegetation indices), and water map layers.
• Methodology: Feature extraction $\rightarrow$ Clustering $\rightarrow$ Risk surface modeling.
• Models: Random Forest on raster data or Spatial Gaussian Processes.
• Ethics: Privacy; ensure maps don’t track individual herds or facilitate stock theft.

9. Abattoir Throughput and Backlog Prediction

• Topic: Supply-Chain Resilience.
• Rationale: Quarantines create massive backlogs at “clean” abattoirs, hurting meat quality and farmer income.
• Objective: Predict the 14-day backlog at specific facilities.
• Data: Historic intake, current transport restrictions, and staffing levels.
• Methodology: Forecast arrivals $\rightarrow$ Queueing theory simulation $\rightarrow$ Scheduling optimization.
• Models: LSTMs for arrival forecasting + Heuristic scheduling.
• Ethics: Labor rights; ensure optimization doesn’t lead to unsafe work speeds for abattoir staff.

10. Pricing + Food Security Impact Model (Beef/Dairy)

• Topic: Econometric ML.
• Rationale: Outbreaks often lead to “Panic Pricing” in supermarkets.
• Objective: Quantify how much of a price spike is due to FMD vs. general inflation.
• Data: Retail price indices, export/import volumes, and outbreak dates.
• Methodology: Causal Inference (Difference-in-Differences) $\rightarrow$ Driver analysis.
• Models: Bayesian Additive Regression Trees (BART).
• Ethics: Accurate attribution; avoid creating models that could be used by retailers to justify price gouging.

11. “Best Inspection Route” for Rapid Response Teams

• Topic: Logistics Optimization.
• Rationale: In provinces like Limpopo, 30 vets may be responsible for hundreds of square kilometers.
• Objective: Minimize travel time for inspection teams while prioritizing high-risk suspected cases.
• Data: Case queue, road network conditions, and team locations.
• Methodology: Risk-weighted Vehicle Routing Problem (VRP).
• Models: Genetic Algorithms or Google OR-Tools.
• Ethics: Geographic fairness; don’t let the algorithm permanently “skip” the most remote, low-tech communities.

12. Crowdsourced Reporting App with ML Triage

• Topic: Participatory Epidemiology.
• Rationale: Farmers see symptoms days before a vet arrives.
• Objective: Provide a platform for farmers to report “suspicious” cases with credibility scoring.
• Data: User reports, GPS (coarse), and metadata.
• Methodology: Triage model $\rightarrow$ Verification loop with local Animal Health Technicians.
• Models: Bayesian Credibility Scoring (ranking reports by historical accuracy).
• Ethics: Data sovereignty; farmers must own their data and consent to sharing it with authorities.

13. Biosecurity Compliance Recommender for Farms

• Topic: Personalized Recommender Systems.
• Rationale: Generic advice is ignored. Farmers need to know the one thing they should fix first.
• Objective: Recommend the top 3 high-impact biosecurity changes for a specific farm.
• Data: Survey of current practices vs. local outbreak outcomes.
• Methodology: Collaborative filtering (farms like yours…) or Uplift modeling.
• Models: Neural Collaborative Filtering or CausalML.
• Ethics: Avoid “Blame Culture”; focus on supportive recommendations rather than punitive audits.

14. Traceability Prototype with Privacy-Preserving IDs

• Topic: Privacy-Enhancing Tech (PETs).
• Rationale: Farmers fear that traceability leads to surveillance and tax targeting.
• Objective: Create a movement ledger that proves an animal came from a “clean” zone without revealing the farmer’s private ID.
• Data: Movement events and district-level location IDs.
• Methodology: Zero-Knowledge Proofs or Differential Privacy.
• Models: Cryptographic protocols (not model-heavy).
• Ethics: Privacy-by-design as the core principle.

15. Outbreak Policy Effect Estimator

• Topic: Policy Evaluation AI.
• Rationale: We need to know: Did the 2025 KwaZulu-Natal lockdown actually work?
• Objective: Statistically estimate the “cases averted” by specific policy decisions.
• Data: Policy timelines across different provinces vs. case growth rates.
• Methodology: Synthetic Control Method.
• Models: Generalized Synthetic Control (GSC).
• Ethics: Guard against political misuse; conclusions must be shared with transparent uncertainty intervals.

16. Multi-Species Risk Model (Cattle, Pigs, and Sheep)

• Topic: One-Health Modeling.
• Rationale: FMD affects various species differently; pigs are “amplifiers,” while sheep often show few symptoms.
• Objective: Predict inter-species spillover risk on mixed farms.
• Data: Farm census (species counts) and regional infection data.
• Methodology: Multi-task learning (sharing features across species).
• Models: Hierarchical Bayesian Models.
• Ethics: Protect proprietary commercial secrets of mixed-farming operations.

17. Synthetic Outbreak Data Generator

• Topic: Generative AI for Science.
• Rationale: Real outbreak data is too sensitive to share with all students/researchers.
• Objective: Create a “digital twin” dataset of an FMD outbreak for training models.
• Data: Published parameters of FMD spread and aggregated stats.
• Methodology: Generative Adversarial Networks (GANs) for time-series.
• Models: TimeGAN or CTGAN (for tabular data).
• Ethics: Ensure the data is truly “synthetic” and cannot be reverse-engineered to identify real farms.

18. Explainable “Farm Risk Scorecard” for Extension Officers

• Topic: Explainable AI (XAI).
• Rationale: If a vet tells a farmer “you are high risk,” they need to explain why to get buy-in.
• Objective: Generate a risk score with plain-language explanations of the drivers.
• Data: Movement history, neighbor status, and biosecurity scores.
• Methodology: Train complex model $\rightarrow$ Extract local explanations.
• Models: SHAP (SHapley Additive exPlanations) or LIME.
• Ethics: Interpretability; ensure the explanation is understandable to a non-technical farmer.

19. Forecasting Export Disruption Risk

• Topic: Trade Intelligence.
• Rationale: Industry bodies need to warn exporters 3 months ahead if a “Red Zone” is expanding.
• Objective: Predict the probability of a province losing its export-free status.
• Data: Case escalation speed and international trade protocol thresholds.
• Methodology: Survival Analysis for “Time-to-Event” prediction.
• Models: Cox Proportional Hazards or DeepHit.
• Ethics: Prevent market manipulation, ensure findings are released to all stakeholders simultaneously.

20. Community Engagement Optimiser for Outbreak Communication

• Topic: Behavioral Analytics & NLP.
• Rationale: If people don’t trust the message, they won’t report cases.
• Objective: Determine which message format (SMS vs. WhatsApp, Language X vs. Y, Tone A vs. B) drives the most reports.
• Data: Engagement metrics from different campaign variants.
• Methodology: Uplift modeling on A/B testing data.
• Models: Causal Forests.
• Ethics: Cultural sensitivity; messages must be factually accurate and respectful of local languages and traditions.

Summary Table for Fast Reference

The FMD crisis is a challenge, but for the next generation of Southern African data scientists, it is a chance to build tools that protect our livelihoods. Which of these will you build?

This is a robust topic for a South African university context. It blends high-stakes veterinary science with cutting-edge AI (uplift modeling) and localized socio-linguistics.

Here is your ready-to-submit mini-proposal template, formatted for clarity and scannability.

Project Proposal: Community Engagement Optimiser for Foot-and-Mouth (FMD) Outbreak

1. Title

Optimising Community Engagement Messaging for Foot-and-Mouth Disease (FMD) Control in South Africa: A Data-Driven, Ethical A/B Testing and Uplift Modelling Approach

2. Background and Problem Statement

Foot-and-mouth disease (FMD) outbreaks in South Africa pose a severe threat to national food security and international trade. Rapid containment relies on community compliance with biosecurity measures: limiting animal movement, early symptom reporting, and adhering to quarantine.

The Problem: Current communication strategies are often “one-size-fits-all” and fail to account for South Africa’s unique socio-technical landscape, characterized by language diversity, a significant digital divide, and varying levels of institutional trust. Consequently, even scientifically sound guidance often fails to trigger the necessary behavioural changes in communal and informal farming sectors. There is currently no systematic, data-driven framework to measure which message variants actually drive compliance across different demographic segments.

3. Literature Review Structure

A comprehensive review will be organized around the following pillars:

• FMD Control in the SA Context: Reviewing the role of the Department of Agriculture, Land Reform and Rural Development (DALRRD) and the impact of informal livestock movements.
• Risk Communication & Behavioural Economics: Exploring the COM-B model (Capability, Opportunity, Motivation, and Behaviour) and the use of “nudges” in public health.
• Digital Divide & Accessibility: Analyzing the penetration of WhatsApp vs. SMS in rural Limpopo, KZN, and Mpumalanga.
• Machine Learning for Social Good: Specifically, Uplift Modelling (Heterogeneous Treatment Effects) to predict individual-level responses to interventions.
• Ethical Data Governance: Reviewing the Protection of Personal Information Act (POPIA) within the context of public health surveillance.

4. Data Requirements

To build a functional optimiser, the project requires:

• The Message Library: 20–40 variants categorized by tone (authoritative/supportive), framing (loss/gain), and language (isiZulu, English, Afrikaans, Sepedi).
• Engagement Logs: Timestamps, channel ID (WhatsApp/SMS), and message ID.
• Outcome Data: * Primary: Reporting rates (USSD/Hotline calls).
  • Secondary: Link click-through rates, keyword responses, or pulse survey completions.
• Contextual Features: District-level data, language preference, and historical outbreak proximity.

5. Methods (Research Design & Workflow)

The research follows a causal inference pipeline:

5.1 Research Design

We will employ a Quasi-Experimental Design (or randomized A/B testing if a pilot is granted). We aim to estimate the Individual Treatment Effect (ITE), defined as:

$\tau(x) = E[Y | X=x, W=1] – E[Y | X=x, W=0]$

Where:

• $Y$ is the behavioural outcome (for example, reporting a case).
• W is the message variant (treatment).
• $X$ represents the community context features.

5.2 Technical Pipeline

1. Preprocessing: Standardizing delivery logs and outcome counts.
2. Feature Engineering: NLP-based features for messages (sentiment, length) and demographic proxies.
3. Modeling: Implementing Causal Forests or X-Learners to identify which segments respond best to which “nudges.”
4. Optimization: A recommendation engine that selects the optimal message variant for a given district/channel.

6. Evaluation Plan

• Statistical Metric: The “Uplift Score” or Qini Coefficient to measure how much better the model performs compared to random message assignment.
• Behavioral Metric: Percentage increase in “High-Quality Reports” (verified cases) vs. “Noise.”
• Practicality: A “Model Card” detailing explainability, why a specific tone was recommended for a specific region.

7. Ethical Considerations

• POPIA Compliance: All data will be de-identified. We will use coarse location data (Ward/District) rather than specific farm coordinates.
• Non-Deception: All messages will remain factually accurate; “testing” only applies to the framing and delivery channel.
• Digital Equity: The project explicitly includes low-tech channels (SMS/Radio scripts) to ensure that the “optimisation” does not inherently exclude farmers with limited internet access.

8. 6–12 Week Deliverable Plan