TerraGuard | Devpost

💥 - How it all started

With climate change increasing the frequency and intensity of natural disasters, communities face escalating risks from wildfires, earthquakes, floods, and hurricanes. Our team developed TerraguardAI to provide accurate risk assessments and actionable insights, empowering government agencies, stakeholders, and at-risk families to make informed decisions, earlier.

📖 - What it does

TerraguardAI is a multi-disaster risk assessment tool with live satellite integration, interactive map, and ML inference to identify at-risk regions enabling governments, individuals, and agencies to take earlier action.

Risk Factor Prediction: Our model utilized a pre-trained and inference-optimized XGBoost classification model to predict likelihood of wildfire, flood, earthquake, and hurricane occurrence in a 1km x 1km grid. We trained our model on historical T + week and T + year data, leveraging a suite of over 60 features including temperature, NDVI (vegation index), wind speed, and air pressure.
Live Satellite Integration: We collect satellite imagery and remotely sensed data from NASA satellites like MODIS and LANDIS, utilizing Google Earth’s API endpoint and pyramid feature extraction to obtain real-time data for various environmental features. This is then fed into our model to generate an interactive, constantly-updated map that shows the most endangered regions.
Agentverse: In order to integrate and combine the large amount of remote sensing data and various predictions, we utilize an agent verse framework that interacts with the user. The “brain” of the agent verse is the manager, which is responsible for aggregating remote sensing data and redistribution to individual agents. Each individual agent is responsible and specialized in a particular natural disaster, communicating with the manager back and forth to generate the best response.
Stakeholder Insights: Stakeholder Insights provides tailored, data-driven recommendations using LLM agents to help governments, insurance companies, and individuals make informed decisions. Governments receive strategies for disaster preparedness and resource allocation, insurers get risk evaluations for adjusting premiums, and individuals receive personalized risk assessments and preventive suggestions. This ensures proactive disaster management and safety measures for all stakeholders.

🔧 - How we built it

Fetch AI Agents: Framework for communicating and coordinating suite of agents
Skylo: live mapping capabilities to incorporate IoT sensor data from satellites efficiently and in real-time
Gemini: Fast, Efficient, Reliable LLM inference endpoint
SKLearn and XGBoost: disaster classification using satellite imagery features
Folium: mapping
Agentverse Framework: Coordinating specialized agents for wildfires, hurricanes, earthquakes, and floods to optimize performance
Flask
TailwindCSS
Axios: Request client to fetch data from that back end to display to the user

📒 - The Efficacy of our Models

Our multi-agent system integrates diverse data sources to enhance prediction accuracy across multiple disaster types. By combining satellite data with advanced machine learning techniques, TerraguardAI provides reliable risk assessments supporting proactive disaster management.

We collected over 16k individual data points for over 60 features in the time period 2010 - 2023. This data was cleaned, standardized, and processed for training using SKLearn. We tested ten diverse models for binary classification using cross validation and found XGBoost performed best.

Model Performance Comparision

To optimize performance, we utilized K-Fold validation for hyperparameter tuning, and feature importance methods like SKLearn’s feature permutation to prune out unnecessary features to optimize inference and performance for specific natural disasters. Ultimately, we achieved over 93% accuracy and directly integrated our model into the Flask backend. This high accuracy speaks to the reliability of our model. However, it's essential to remain vigilant against overfitting and conduct thorough validation to ensure its generalizability, a testament to our commitment to both performance and robustness.

First gradient boosting tree

🚩 - Challenges we ran into

Real-time data collection and model inference Communication efficiency for integrating front and backend Displaying map UI elements and heatmap Developing a user-friendly interface for complex data visualization.

🏆 - Accomplishments that we're proud of

Successfully creating a comprehensive multi-disaster risk assessment tool.
Success development and integration of Agentverse framework to provide personalized stakeholder recommendations.
Collaboration among teammates
Real-time updates

📝 - What we learned

Internals of function decorators in Python and how they are utilized for Fetch AI. We learned the importance of data integration and machine learning in addressing complex environmental challenges. Our experience highlighted the value of collaboration and innovation in developing effective solutions.

✈️ - What's next for TerraguardAI

Phase 1: Expansion

Enhance predictive models with additional environmental parameters.
Increase collaboration with government agencies and insurance firms.

Phase 2: Mobile Integration

Develop mobile applications for broader accessibility.
Expand features to cover additional natural disasters.

Phase 3: Global Reach

Localize services for international markets.
Partner with global organizations to enhance disaster resilience.

📋 - Evaluator's Guide to TerraguardAI

We utilized cutting-edge technologies such as Google Earth Engine, machine learning models, and LLMs to develop TerraguardAI. Our platform is designed to provide comprehensive insights into multiple disaster risks, empowering stakeholders to make informed decisions and improve community resilience.

Built With

Fetch.ai for AI Agents We designed a LLM hierarchy framework based on a central controller responsible for communciating between LLM agents and main application, and individual LLM agents fine-tuned for specializaed recommendations for specific disasters. These agents are designed to fetch data from various related resources based on the user's needs, and utilize LLMs themselves to address problems like which agent to route to. This allows for seamless data retrieval and integration, ensuring that users receive the most relevant and up-to-date information. It also serves as a form of test-time scaling that improves the relevance, accuracy, and effectiveness of our recommendations. The AI agents are crucial for automating data collection and enhancing the overall efficiency of our platform.

Gemini for LLM Models Our agents are backed by the power of Gemini, a cutting-edge framework for Large Language Models (LLMs). We utilized Gemini Turbo 1.5, which was essential as we needed speed and long context to communicate between the various agents. Gemini is also trained on vast internet data, enabling our agents to give relevant answers to user questions. The logical reasoning ability of the Gemini framework is also extensively utilized, we utilize the LLM internally as a decision-maker to decide what needs to be communicated between controller and workers and who to communicate to, specific parameters to query the main application, model routing, and more. By leveraging Gemini, we can provide important information about properties and their longevity, helping users make informed decisions. Gemini's advanced capabilities ensure that our agents can deliver accurate and insightful responses, enhancing the user experience.

Skylo for Live Mapping Skylo is another integral part of our platform, providing live mapping capabilities. Skylo's technology allows us to deliver real-time updates and visualizations of natural disasters and other relevant geospatial data. This live mapping feature is essential for users who need immediate and accurate information about ongoing events. By integrating Skylo, we can offer a dynamic and responsive mapping experience, ensuring that users are always informed about the latest developments. By combining these powerful technologies, we create a comprehensive and robust platform that delivers real-time, accurate, and actionable information about natural disasters and related events. This integration ensures that our users have access to the most advanced tools and insights, enabling them to make informed decisions and stay safe.