Land Change Detection

Local Streamlit application for visual land-surface change analysis on paired satellite images.

The current runnable app uses:

• mfaytin/mask2former-satellite for RGB surface segmentation diagnostics.
• gemma4:e4b through Ollama for the final natural-language visual interpretation.
• OSCD sample imagery by default, with optional user-uploaded image pairs.

The normal user-facing report is VLM-first: Gemma receives only the before crop, the after crop, and a labeled A1..D4 before/after contact sheet. It does not receive Mask2Former maps, class labels, transition tables, DINO features, or heatmaps.

What Works Now

The Streamlit UI currently provides:

• before/after crop selection;
• Mask2Former surface segmentation diagnostics from artifacts/models/semantic/mask2former-satellite;
• a labeled 4x4 before/after comparison grid;
• Gemma visual interpretation with:
  • scene overview;
  • before summary;
  • after summary;
  • main visible changes;
  • per-cell observations for A1..D4.

Runtime Models

Required for the current app:

Research or debug only:

Requirements

Supported development targets:

• macOS, Linux, or Windows.
• Python 3.11 recommended.
• Git.
• For VLM output: Ollama installed and running.

Hardware notes:

• CPU works, but model inference is slower.
• Apple Silicon can use PyTorch MPS for Mask2Former.
• NVIDIA CUDA can be selected in the UI when PyTorch/CUDA is installed correctly.
• Ollama manages Gemma execution separately from PyTorch device selection.

Setup On macOS / Linux

git clone https://github.com/SargisVardanian/Land-Change-Detection.git
cd Land-Change-Detection

chmod +x scripts/create_env.sh
./scripts/create_env.sh
source .venv/bin/activate

If your system does not have python3.11, install Python 3.11 first, or run:

PYTHON_BIN=python3 ./scripts/create_env.sh

Setup On Windows

Use PowerShell:

git clone https://github.com/SargisVardanian/Land-Change-Detection.git
cd Land-Change-Detection

py -3.11 -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install --upgrade pip setuptools wheel
python -m pip install -e ".[dev]"

If PowerShell blocks activation, run:

Set-ExecutionPolicy -Scope CurrentUser RemoteSigned

Then activate the environment again.

Download The Required Segmentation Model

The current app needs only Mask2Former:

source .venv/bin/activate
python scripts/download_semantic_models.py

This downloads:

artifacts/models/semantic/mask2former-satellite

Optional research checkpoints are not required for the app. Download them only if you need Prithvi experiments:

python scripts/download_semantic_models.py --include-research

Install And Prepare Ollama

Gemma runs through Ollama, which is a separate local application/server.

Install Ollama:

• macOS: download from ollama.com/download
• Windows: download from ollama.com/download/windows
• Linux:

curl -fsSL https://ollama.com/install.sh | sh

Start Ollama if it is not already running:

ollama serve

In another terminal, pull the Gemma model used by the app:

ollama pull gemma4:e4b

Check that Ollama sees it:

ollama list

Expected entry:

gemma4:e4b

If your Ollama registry does not provide gemma4:e4b, install an equivalent local vision-capable Gemma model and update GEMMA4_E4B_OLLAMA in src/land_change_detection/remote_sensing_vlm.py.

Download Sample Data

The app can run with OSCD sample imagery:

source .venv/bin/activate
python scripts/download_oscd.py

You can also use your own before/after image pair from the UI.

Run The App

source .venv/bin/activate
streamlit run app.py

Open the URL printed by Streamlit, usually:

http://localhost:8501

Recommended first run:

1. Keep Image source as OSCD dataset.
2. Keep Semantic model as Mask2Former satellite / OpenEarthMap classes.
3. Keep Available model as gemma4:e4b (Ollama).
4. Keep Reasoning budget as Full local analysis.
5. Select or confirm a crop.
6. Wait for the Gemma visual interpretation.

Normal UI Contract

Normal mode:

• shows selected before/after crop;
• shows Mask2Former segmentation diagnostics;
• shows the A1..D4 visual comparison grid;
• sends only RGB before/after/contact-sheet images to Gemma;
• shows Gemma's final English report.

Debug mode:

• may show raw prompts, raw model output, runtime details, and semantic diagnostics;
• does not change the normal Gemma input contract.

Experimental/heavy VLMs:

• hidden by default;
• intended only for research/debug;
• not recommended for normal local use.

Why DINOv3 Is Not Used As Segmentation

DINOv3 SAT models such as facebook/dinov3-vitl16-pretrain-sat493m and timm/vit_large_patch16_dinov3.sat493m are feature-extraction backbones. They do not include a trained land-cover segmentation decoder/head in this project, so they cannot directly output classes such as road, building, water, or bare land.

For that reason, DINOv3 is not part of the Streamlit runtime.

Architecture Direction

The longer-term research architecture remains semantic-first:

T1 semantic segmentation -> T2 semantic segmentation -> transition matrix -> interpretable report

Planned next steps:

1. Keep Mask2Former as the current RGB baseline.
2. Integrate Prithvi-EO-2.0-300M-TL through a proper multispectral TerraTorch path.
3. Add Prithvi-EO-2.0-600M-TL only after the 300M path is stable.
4. Add CDMamba as a binary changed/unchanged validation baseline, not as the main semantic answer.
5. Keep VLMs as explanation/reporting tools, not as pixel-mask evidence generators.

Verification

Run static and unit checks:

source .venv/bin/activate
PYTHONPATH=src python -m compileall -q app.py src tests
PYTHONPATH=src python -m pytest -q

Expected current result:

26 passed

Sources

• mfaytin/mask2former-satellite: Hugging Face
• Prithvi-EO-2.0-300M-TL: Hugging Face
• Prithvi-EO-2.0-600M-TL: Hugging Face
• Prithvi-EO-2.0 paper: arXiv:2412.02732
• CDMamba: arXiv:2406.04207
• EarthDial: arXiv:2412.15190
• Ollama installation: ollama.com/download

Land-Change-Detection