🌳 Sardinia Forest Dieback EO

Which forest stands should field teams inspect first after the 2024 drought? Sentinel-2 time series, geospatial foundation models and an MCP agent powered by EVE.

From 98 Sentinel-2 acquisitions to a ranked inspection list (illustrative render).

Table of contents

• Overview

• Highlights

• Quick start

• Results

• How it works

• Foundation models (TerraMind, TESSERA)

• Agent: the pipeline as MCP tools, called by EVE

• Project structure

• Reproducibility

• Limitations

• References and acknowledgements

• License

Related MCP server: GeoSight MCP Server

Overview

In August 2024 the Sardinian Regional Government declared urgent measures against a dieback of cork-oak and holm-oak woods and Mediterranean scrub, spreading over thousands of hectares, and charged the regional forest agency FoReSTAS with monitoring and response (Delibera RAS n. 31/7, 28.08.2024).

This project answers the question a forest manager has to settle before sending teams out: which forest areas in the Sette Fratelli massif (Sarrabus, SE Sardinia) dried out in summer 2024 and did not recover in 2025? In a drought year almost every crown loses water. The stands that fail to recover the following year are the ones at risk of dieback and mortality.

Answer in brief. Summer 2024 was by far the driest canopy state of 2019–2025. 2,211 ha (13 % of the analysed forest) dried out much more than the surrounding woods and had not recovered by summer 2025, after excluding fires and clear-cuts. The alert is densest in 40 blocks of 25 ha where at least half of the forest is flagged (528.6 ha, about a quarter of the total), ranked in outputs/alert_cells.csv / .geojson.

Everything is in notebook.ipynb, narrated and executed top to bottom.

Highlights

• 🛰️ Real pixels, no account: 98 Sentinel-2 L2A acquisitions (July–September 2019–2025) streamed from Planetary Computer via STAC.

• 🧹 Careful preprocessing: processing-baseline offset correction, SCL cloud/shadow whitelist, per-summer median composites on the native 20 m UTM grid.

• 💧 Decision rule: a landscape-relative NDMI anomaly in 2024 and less than 50 % recovery in 2025. Fires and logging are screened out with NBR step changes. Threshold sensitivity is reported.

• 🗺️ Decision-ready outputs: ranked 25 ha cells (CSV, GeoJSON), Cloud-Optimised GeoTIFFs, an interactive folium map, and a plain-language summary generated from the numbers.

• 🧠 Foundation models: TerraMind 1.0 run locally (generated land cover, embeddings) and TESSERA embeddings add what band math cannot tell.

• 🤖 Agentic: the same pipeline exposed as MCP tools and called by EVE-Instruct, running locally, through a LangGraph agent. Each answer is checked automatically against the tool outputs.

• ✅ Tested: pytest runs the real MCP server and the agent wiring in CI, no LLM required.

Quick start

Requires Python 3.12 with the venv module (on Ubuntu/Debian: sudo apt install python3.12-venv, or create the environment with pyenv or uv).

git clone https://github.com/bitwiseops/sardinia-forest-dieback-eo.git
cd sardinia-forest-dieback-eo
python3.12 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

jupyter lab notebook.ipynb                     # the full analysis (mostly downloads, so time depends on your network)
pytest -q                                      # MCP server + agent wiring, no LLM needed

Optional, to run the agent with EVE (14 GB model). It needs a recent Ollama (tested with 0.40.6): with older releases (tested: 0.23.0) the EVE model is registered without tool support and the agent stops with does not support tools. Check with ollama --version.

ollama pull hf.co/eve-esa/EVE-Instruct-GGUF-Q4_K_M
python -m mcp_server.agent --live "After the 2024 drought, which forest areas in bbox [9.33, 39.19, 9.52, 39.33] should the forest service inspect first?"
python -m mcp_server.agent --model qwen2.5:7b-instruct "..."    # any Ollama model with tool calling

The repository ships the tool results of the example question in outputs/mcp/cache/, so that first command is answered from cache (the tool result says "cached": true). To run the pipeline on fresh Sentinel-2 data, delete the cache first: rm -r outputs/mcp/cache.

Results

The three top-ranked cells in real Sentinel-2 summer composites: green in 2023, brown in 2024, still grey-brown in 2025 while the surrounding forest has turned green again.

How it works

1. Ingest: STAC search on Planetary Computer, tile-level cloud pre-filter, one item per date (newest processing baseline).

2. Preprocess: DN to reflectance with the +1000 offset removed (baseline ≥ 04.00), SCL whitelist (vegetation, not-vegetated), 20 m UTM 32N grid, July–September median per year.

3. Indices per acquisition: NDMI (canopy water, main), NDVI, CRswir (continuum removal in the SWIR), NBR (disturbance screen).

4. Anomaly and recovery: subtract each year's forest-wide median (removes the drought shared by the whole area), then per pixel z = (x₂₀₂₄ − μ)/σ over 2019–2023 and R = share of the drop regained in 2025, following the resistance/recovery framework of tree-ring ecology (Lloret et al., 2011) as applied to Sentinel-2 by Sturm et al. (2022).

5. Decide: alert if z ≤ −2 and R < 0.5, aggregate on 500 m cells, rank by flagged hectares.

All functions live in src/pipeline.py: the notebook and the MCP server call the same code.

Foundation models (TerraMind, TESSERA)

TerraMind 1.0 base (IBM, ESA, Forschungszentrum Jülich) runs locally on 12-band 10 m composites. Its generated land cover agrees with the real ESRI / IO 2023 map on roughly 89 % of pixels (the exact value changes slightly between runs). See notebook section 9 and src/gfm.py.

Agent: the pipeline as MCP tools, called by EVE

Real terminal recording: EVE-Instruct, running locally, answers the example question of the assignment by calling compare_vegetation. The pipeline runs on fresh Sentinel-2 data and the answer passes the grounding check.

• Server (mcp_server/server.py, FastMCP, stdio):

  • three tools: list_scenes, compare_vegetation, dieback_alerts;

  • validated inputs, with errors returned as readable messages;

  • results cached by arguments.

• Agent (mcp_server/agent.py):

  • a LangGraph ReAct graph with langchain-mcp-adapters;

  • EVE-Instruct (eve-esa/EVE-Instruct-GGUF-Q4_K_M) served by Ollama.

• Transcripts (outputs/mcp/):

  • model digest, exact tool calls, raw results, answer, and an automatic grounding check that every number in the answer comes from a tool;

  • earlier runs are kept in run1/ and run2/, including the invented number the check caught.

Project structure

├── notebook.ipynb            # narrated analysis, all cells executed
├── src/
│   ├── pipeline.py           # STAC search, masking, indices, composites, anomaly/recovery, cells, export
│   └── gfm.py                # TerraMind (local) and TESSERA embeddings
├── mcp_server/
│   ├── server.py             # MCP server: list_scenes, compare_vegetation, dieback_alerts
│   ├── agent.py              # LangGraph agent + EVE via Ollama, transcripts, grounding check
│   └── demo_questions.py     # the three demonstration questions
├── tests/                    # pytest: real MCP server, agent wiring with a scripted fake LLM
├── outputs/                  # generated rasters, vectors, figures, MCP transcripts
├── figures/                  # method diagrams and flow figure, terminal recording
├── requirements.txt          # pinned versions used to execute the notebook
└── requirements-ci.txt       # lighter set for CI (no foundation models)

Reproducibility

• Environment: Python 3.12, versions pinned in requirements.txt. Executed on macOS arm64 (CPU only), and re-run from a clean clone on Ubuntu 24.04 x86_64 (notebook and tests on CPU).

• Run time: the notebook is mostly downloads, so it depends on your network and CPU. TESSERA embeddings are cached in data/ on first run (several GB of disk).

• What is exactly reproducible: the band-math results (analysed and flagged forest area, priority cells and their values) were identical between the original run and the clean-clone run. Values generated by TerraMind (for example canopy_no_longer_trees and the land-cover agreement) change slightly from run to run, and cells with the same flagged area can swap rank.

• No credentials: Planetary Computer signs asset URLs automatically.

• EVE is optional: without Ollama, notebook section 10 shows the transcripts of the last live run and says so.

Limitations

There is no field validation: alerts are hypotheses for field checks, not validated dieback maps. Only one post-drought year (2025) is available, so slow resprouting can look like non-recovery. One forest class (WorldCover tree cover) covers several species. No topographic correction is applied. The full discussion, with trade-offs and next steps, is in notebook section 8.

References and acknowledgements

The key references are Lloret et al. (2011), Oikos; DeSoto et al. (2020), Nature Communications; Sturm et al. (2022), Global Change Biology; Mouret et al. (2024), IEEE JSTARS. The full list with DOIs is at the end of the notebook.

Data and models: Copernicus Sentinel-2 (ESA), Microsoft Planetary Computer, ESA WorldCover 2021, ESRI / Impact Observatory land cover, TerraMind (IBM, ESA, Forschungszentrum Jülich), TESSERA, and EVE (ESA Φ-lab, Pi School).

Developed with AI coding assistance, as the assignment allows. The use case, the methodological choices and the results were reviewed by me, and I will explain them in the interview.

License

MIT © 2026 Flavio Cordari