AI Climate Resilience Sandbox

An interactive scenario-planning tool that helps communities explore climate risks, test adaptation strategies, and generate resilience guidance for drought and water-stress conditions.

Comment

About The Project

Inspiration

This project was inspired by the difficulty local decision-makers face when trying to prepare for climate stress before a crisis occurs. ENSO events such as El Niño can contribute to rainfall deficits, heat stress, drought conditions, crop losses, and water shortages. While many climate tools focus on monitoring or forecasting, there is also value in a simple planning tool that lets users explore “what-if” scenarios and compare possible adaptation responses.

What It Does

AI Climate Resilience Sandbox is an interactive Streamlit app for exploring hypothetical climate-risk scenarios for Tamil Nadu. Users can adjust sea surface temperature anomaly, monsoon rainfall deficit, drought duration, and heat stress, then toggle adaptation measures such as drought-resistant crops, groundwater rationing, supplemental irrigation, and water conservation.

The app displays planning indicators including crop yield stability, water reservoir security, groundwater stress, agricultural risk exposure, and a regional resilience score. It also includes a risk map, z-score-based risk routing, policy recommendation cards, transparency notes, and exportable scenario briefings.

How We Built It

The project was built with Python, Streamlit, Plotly, Pandas, NumPy, scikit-learn, and joblib. We used a historical reference dataset for Tamil Nadu covering 1991-2017, combining climate, rainfall, crop, and reservoir-related indicators.

We trained Random Forest surrogate models to estimate crop and water-security outcomes. The app then combines those outputs with rule-based calculations for groundwater stress, agricultural risk, resilience scoring, and policy recommendation routing. Policy guidance and thresholds are stored separately in JSON files so they can be reviewed or updated without changing the core app code.

Challenges We Faced

The biggest limitation was the small dataset size. The model is trained on 27 yearly records, which is not enough for high-confidence forecasting or fine-grained district-level prediction. Because of this, the model should be treated as a scenario-planning prototype, not a forecasting system.

Another challenge was making the adaptation comparison meaningful. Random Forest models can produce flat or plateaued outputs when new scenarios fall into the same learned decision leaves, especially with small datasets. To make the app usable for planning, we added simplified adaptation-impact assumptions to the displayed indicators. These assumptions are transparent but should not be interpreted as validated causal estimates.

We also faced deployment and usability issues, including dependency setup, Streamlit redeployment behavior, Windows console Unicode issues, and making the dashboard readable for non-technical users.

What We Learned

We learned that climate decision-support tools need to communicate uncertainty clearly. A model output alone is not enough; users need to understand the historical data range, assumptions, confidence limits, and cases where a scenario exceeds the training data.

We also learned the value of separating the model layer from the policy layer. Keeping recommendations, metadata, and thresholds in configuration files makes the system easier to inspect and update.

Limitations

This project has important limitations:

  • It is not an operational forecast.
  • It does not predict future ENSO events or actual drought declarations.
  • The dataset is small and historical coverage is limited.
  • Some indicators are simplified proxies rather than direct measurements.
  • Adaptation effects are modeled as planning assumptions, not proven real-world impacts.
  • The regional map is illustrative and should not be treated as validated district-level risk mapping.
  • Policy recommendations are general guidance and do not replace official government advisories.

Future Scope

Future versions could improve reliability by using larger datasets, more granular district-level data, additional crop and groundwater indicators, better model validation, and integration with live climate and reservoir data sources. The adaptation-impact assumptions could also be replaced with calibrated estimates from agronomic, hydrological, or field-level studies.

Built With

Share this project:

Updates