EcoRevive

Inspiration

Every year, wildfires devastate millions of acres across the western United States. Satellite imagery can show us where fires burned — but that's only 10% of the problem. The other 90% is answering: "What do we actually do about it?"

Restoration planners at NGOs spend weeks manually stitching together fragmented data — outdated species guides, scattered regulatory frameworks, incomplete severity assessments — before they can even begin writing a grant proposal. Community volunteers show up wanting to help but have no idea where it's safe to work or what to plant. Meanwhile, ecosystems are on a ticking clock — the first two growing seasons after a fire are critical for preventing erosion, invasive species takeover, and permanent habitat loss.

We built EcoRevive because we believe the gap between detecting damage and taking action shouldn't take weeks. It should take seconds.

What it does

EcoRevive is an end-to-end wildfire restoration intelligence platform. Users click anywhere on an interactive 3D globe, and within seconds receive a complete, science-backed restoration plan — from burn severity mapping to species prescriptions to permit guidance.

Core capabilities:

Burn Severity Mapping — A custom U-Net deep learning model (trained on 7 major California wildfires) processes 10-band Sentinel-2 satellite imagery to produce continuous burn severity maps at 10-meter resolution.
Ecosystem Classification & Species Recommendations — Gemini 3 analyzes both the RGB satellite image and the severity overlay simultaneously to classify the ecosystem, identify the reference vegetation community, and prescribe native species with specific planting densities per hectare.
Safety Hazard Detection — Automatically identifies widowmaker zones (dead standing trees), landslide-prone slopes, and access restrictions so volunteers don't walk into dangerous areas.
Carbon Sequestration Estimates — Calculates how much carbon the burned area lost, how much restoration will recapture, and translates it into relatable terms ("equivalent to taking 847 cars off the road for a year").
Recovery Timeline Forecasting — Projects vegetation recovery at Year 1, 5, 15, and 30, showing expected ground cover, shrub establishment, and canopy closure percentages.
Land Ownership & Permit Lookup — Determines whether the land is federal, state, private, or tribal, identifies the managing agency, and lists required permits for restoration work.
Dual-Persona Reports — The same analysis generates two distinct outputs:
- Community Organizers: A 1-2 page shareable impact card with safety alerts, hope visualizations, and simple action steps.
- Professionals: A 5-10 page grant-ready technical report with IPCC-tier carbon estimates, compliance checklists, and detailed species prescriptions.
Conversational AI Chat — Users can ask follow-up questions about any aspect of the analysis: "What species should I plant on north-facing slopes?", "What are the landslide risks?", "What permits do I need for this federal land?"

How we built it

We designed EcoRevive as a 3-layer architecture with clear separation of concerns:

Layer 1 — Vision Engine (Custom Deep Learning) We trained a U-Net with attention gates on 10-band Sentinel-2 satellite imagery from 7 major California wildfires (Dixie, Caldor, Creek, Camp, Mendocino Complex, Thomas, Rim fires). Training labels came from MTBS and RAVG official burn severity datasets. The model outputs a continuous severity map (0-1) at 256x256 pixels. Kincade and Woolsey fires were held out for validation.

Layer 2 — Reasoning Engine (Gemini 3) This is where EcoRevive leverages six distinct Gemini 3 capabilities:

Multimodal Vision Analysis — We send Gemini two images simultaneously (the RGB satellite tile and the burn severity overlay) alongside spatial context prompts. Gemini reasons across both to identify fragmentation patterns, severity gradients, and restoration-relevant features.
Structured JSON Output — Using response_mime_type='application/json', Gemini returns machine-readable spatial analysis (zone classifications, risk grids, confidence scores) that plugs directly into our frontend and PDF generators.
Function Calling (Agentic Behavior) — Gemini acts as an orchestrating "brain" that autonomously calls specialized tools: analyze_ecosystem(), check_safety_hazards(), lookup_land_ownership(), and forecast_recovery().
Google Search Grounding — Real-time retrieval of current species data, regulatory changes, and ecological research ensures recommendations reflect the latest science.
RAG with Gemini Embeddings — We use text-embedding-004 to build a vector store over California-specific species data, ecoregion info, restoration best practices, and legal guidelines. Retrieved context is injected into prompts for domain-grounded responses.
Multi-Turn Chat Sessions — Extended conversations about analysis results, maintaining context across turns for detailed follow-up questions.

Layer 3 — Output Generator (Context & Reports) Land use classification identifies cautions (urban proximity, agricultural boundaries). A carbon calculator produces both fun-fact equivalencies and IPCC-tier scientific estimates. The PDF engine generates persona-specific reports — shareable impact cards for communities, technical documents for professionals.

Users can chat with Gemini in real time to have any mind of questions answered.

Tech Stack:

Frontend: React 19, Vite 7, Cesium.js (3D globe), Three.js & React-Three-Fiber (landing animations)
Backend: FastAPI, Python 3.11, Google Earth Engine API
ML: PyTorch (U-Net with attention gates)
AI: Gemini 3 Flash Preview via google-genai SDK, text-embedding-004
Reports: ReportLab (PDF), python-docx (Word)
Infra: Docker, Render

Challenges we ran into

Getting clean satellite data was harder than expected. Cloud cover frequently obscured post-fire imagery, forcing us to implement aggressive filtering (< 20% cloud threshold) and flexible date-range windows to find usable Sentinel-2 scenes.

Making the U-Net generalize across fire types. Our model initially overfit to specific vegetation patterns. Adding attention gates and training across 7 diverse fires (from coastal scrub to high-altitude conifer) taught the model to focus on burn-relevant spectral signatures rather than memorizing vegetation types.

Taming Gemini's output for production use. Early prompts produced verbose, unstructured text that broke our frontend. We iterated extensively on — dropping temperature to 0.1, enforcing structured JSON schemas, and designing few-shot examples with explicit spatial reasoning patterns.

Multimodal spatial reasoning was non-obvious. Getting Gemini to meaningfully reason across two images (RGB + severity overlay) required carefully describing what each image represents, what spatial patterns to look for, and how to correlate features between them. This took dozens of iterations to get right.

Preventing ecologically harmful recommendations. Gemini could easily recommend non-native invasive species or suggest activities that violate protected area regulations. We built a multi-layer constraint system: hard-coded rules that reject known invasives, LLM-level system prompts with ecological guardrails, and post-processing validation checks.

Balancing two very different user personas. Community organizers need emotional impact and simplicity. Professionals need scientific rigor and legal compliance. Making the same pipeline serve both without compromising either required careful output architecture.

Accomplishments that we're proud of

True multimodal AI — not just a chatbot. EcoRevive sends actual satellite imagery and deep learning model outputs directly to Gemini for visual reasoning. This creates a genuine computer-vision-to-LLM pipeline where the AI reasons over what it sees, not just what it reads.

A custom deep learning model trained on real wildfire data. We didn't just call an API. We built and trained a U-Net with attention gates from scratch on 10-band multispectral satellite data, using official government burn severity labels.

Responsible AI that actually works. Our multi-layer constraint system prevents non-native species recommendations, flags Indigenous land rights, and respects protected area regulations — without degrading output quality. In fact, constraints improve quality by narrowing the solution space.

From satellite bands to PDF report in one click. The full pipeline — Earth Engine data acquisition, U-Net inference, Gemini multimodal analysis, carbon estimation, safety assessment, and formatted report generation — runs end-to-end without manual intervention.

Six Gemini 3 features working in concert. Multimodal vision, structured output, function calling, search grounding, embeddings-based RAG, and multi-turn chat — all integrated into a single coherent application, not used in isolation.

What we learned

Gemini 3's multimodal capabilities are genuinely powerful for geospatial analysis. It can identify spatial patterns in satellite imagery — fragmentation zones, severity gradients, edge conditions — that would require specialized computer vision algorithms to detect programmatically.

Function calling is transformative. It turns Gemini from a text generator into an agentic coordinator. Each tool (analyze_ecosystem, check_safety_hazards, etc.) is independently testable and swappable, and Gemini decides which to call and in what order based on the analysis context.

Structured JSON output is non-negotiable for production. Free-text LLM responses are too fragile for downstream processing. Enforcing JSON schemas made our entire pipeline reliable and debuggable.

Domain-specific RAG dramatically outperforms zero-shot prompting for ecological recommendations. Generic Gemini knows about ecology broadly, but injecting California-specific species data and ecoregion information made recommendations specific, actionable, and correct.

Constraints improve quality, not just safety. By preventing ecologically invalid outputs (invasive species, impossible planting densities), our guardrails actually help Gemini produce better recommendations — the narrower valid search space leads to more precise answers.

What's next for EcoRevive

Expand Beyond California — Adapt the model and knowledge base to the entire western US, Australia, and Mediterranean fire-prone regions. The architecture is designed to be modular: swap the knowledge base and retrain the U-Net on new fire data.
Multi-Hazard Expansion — Extend the platform beyond wildfire to floods, landslides, deforestation, and any ecosystem degradation event that benefits from satellite-to-action intelligence.
Temporal Recovery Monitoring — Track actual vegetation recovery against our predicted timelines using time-series satellite analysis, creating a feedback loop that improves the model over time.
Drone Imagery Integration — Incorporate high-resolution drone data for sub-meter assessment of individual tree mortality, regeneration, and planting site suitability.
Automated Grant Applications — Auto-generate complete FEMA, Cal Fire, and foundation grant proposals pre-filled with analysis data, carbon estimates, and species prescriptions.
Mobile Field Companion App — A lightweight app for on-the-ground volunteers to navigate to assigned planting zones, validate AI predictions against real conditions, and report restoration progress.