---
base_model:
  - ibm-esa-geospatial/TerraMind-1.0-base
pipeline_tag: image-classification
tags:
  - methane
  - detection
  - geospatial
  - terramind
---

# FAST-EO Use Case 2 - Methane Detection

This repository contains data and code to reproduce experiments for fine-tuning **TerraMind-Base** to detect methane-related signatures in multispectral imagery. It includes multiple experiment variants and their corresponding datasets.

The file `Methane_benchmark_patches_summary_v3.xlsx` provides per-patch descriptions and defines the **fold splits** used to ensure non-overlapping partitions. Runner scripts use this Excel file to build train/val/test splits, typically reserving one fold for testing.

All scripts provide usage instructions via `--help` (or `-h`).

## Requirements

- Install the TerraMind/Terratorch stack used by the project before running experiments.
- Ensure your environment has the required dependencies for data loading and training.

## Experiments

### Experiment 1: Fine-tuning on Methane Benchmark Dataset (MBD)

Fine-tune TerraMind-Base on the Methane Benchmark Dataset. The normalized dataset is provided in:

- `MBD_nan_S2_zscore/`

Training code is located in:

- `classification/` (includes dataset and dataloader classes)

### Experiment 2: Fine-tuning on MBD with text captions

This experiment modifies the TerraMind-Base model to concatenate text-caption embeddings with visual embeddings.

- Caption embeddings are computed with `all-MiniLM-L6-v2`.
- Code and resources are in `classification_with_text/`.
- Original captions: `classification_with_text/MBD_text/`
- Precomputed embeddings: `combined_caption_embeddings.csv`

### Experiment 3: Sentinel-2 with simulated atmospheric conditions

Evaluate generalization on Sentinel-2 data with simulated atmospheric conditions in both:
- Top-of-Atmosphere (TOA)
- Bottom-of-Atmosphere (BOA)

The model can be trained on this simulated data or used only for inference to test cross-domain robustness.

### Experiment 4: Intuition-1 with simulated atmospheric conditions

Analogous to Experiment 3, but using Intuition-1 imagery with simulated atmospheric conditions (TOA and BOA). This experiment tests robustness under domain shift.

### Experiment 5: Urban dataset without methane (false-positive stress test)

A control dataset containing only urban imagery without methane is used to check whether models learn methane-specific cues rather than urban signatures. The goal is to quantify false positives.

Scripts for loading and running inference on this dataset are provided in the repository.

## Notes

- Use `--help` on each runner/training script to see available options.
- Keep fold definitions consistent with `Methane_benchmark_patches_summary_v3.xlsx` to ensure comparable results across experiments.