Fields of the World

1. Motivation¶

You now understand what Mask R-CNN predicts and how. This page covers the data it will train on: a benchmark dataset built specifically for field-boundary instance segmentation, and what you need to check and prepare before training begins on the next page.

Instance-level labels are expensive to produce, which is exactly why a well-curated benchmark dataset is valuable. Using one lets you focus on the modeling workflow itself in this lesson, while still working with realistic, high-quality data, the same reasoning that motivated using NWPU-VHR-10 in L06 and the Chesapeake land-cover data in L07.

2. Core idea¶

The Fields of the World (FTW) dataset pairs Sentinel-2 imagery with instance segmentation masks for agricultural field boundaries, across 24 countries. Each chip already comes with a unique integer ID per field baked into its label, exactly the instance-identifier structure from the previous page, so no manual instance derivation is needed. This lesson uses the Luxembourg subset, one of the smallest country subsets, which keeps download and training time manageable while still using real, high-quality annotations.

3. Workflow¶

The Fields of The World (FTW) dataset contains over 70,000 agricultural image chips across multiple countries. For this workflow, we focus on the Luxembourg subset—a compact dataset ideal for learning instance segmentation pipelines without massive download overhead. Each chip uses 4-band Sentinel-2 imagery (Red, Green, Blue, Near-Infrared) at 10-meter spatial resolution. A $256 \times 256$ pixel tile covers a physical area of approximately $2.56\text{ km} \times 2.56\text{ km}$, paired with instance labels marking individual agricultural field boundaries.

A. Download the Luxembourg subset¶

import geoai

geoai.download_ftw(countries=["luxembourg"], output_dir="ftw_data")

The download_ftw() function retrieves the Luxembourg subset and extracts it into the ftw_data/ directory.

B. Explore metadata and spatial splits¶

The dataset includes a GeoParquet file recording every chip’s location, geometric footprint, and official dataset split (train, val, test).

from pathlib import Path
import geopandas as gpd

country_dir = Path("ftw_data") / "luxembourg"
chips_gdf = gpd.read_parquet(country_dir / "chips_luxembourg.parquet")

print(f"Total chips: {len(chips_gdf)}")
print(chips_gdf["split"].value_counts())

geoai.view_vector_interactive(chips_gdf, column="split")

Visualizing data splits on an interactive map is a critical step in geospatial machine learning. A sound benchmark spreads splits evenly across different geographic landscapes. If test chips cluster in a single region, evaluation results will be biased toward that specific terrain or crop type rather than testing true Model Generalization.

C. Preview image-mask pairs¶

geoai.display_ftw_samples("ftw_data", country="luxembourg", num_samples=4)

Each sample pair displays the Sentinel-2 imagery next to its instance mask. In the mask, each distinct color represents an individual field instance (e.g., $1, 2, 3, \dots$), preserving unique object identities rather than simply labeling all agricultural pixels as a single class.

When inspecting samples, perform these manual quality checks:

• Are field boundaries clearly distinguishable in the multi-band imagery?

• Does every instance mask match a plausible agricultural field?

• Do neighboring parcels maintain distinct IDs, or do they merge?

D. Prepare data for training¶

Before passing data to the Mask R-CNN training pipeline, prepare_ftw() handles data rescaling and directory structuring:

data = geoai.prepare_ftw("ftw_data", country="luxembourg")
data

The preparation pipeline executes four main tasks:

raw FTW imagery (16-bit reflectance, 0-10'000)
                   ↓
   rescale values to 8-bit (0-255 uint8)
                   ↓
organize into standard directory layout (images/ + labels/)
                   ↓
   pool dev chips & reserve test set for inference

1. Reflectance Rescaling: Converts raw Sentinel-2 16-bit reflectance values (0 - 10’000) to the 8-bit uint8 range (0 - 255) expected by the underlying geoai Mask R-CNN framework. (Note: This conversion is a specific pipeline requirement for this teaching workflow, not a general rule for all remote sensing tasks).

2. Directory Formatting: Places paired GeoTIFF tiles into standardized subfolders:

field_boundaries/
├── images/
└── labels/

3. Split Reorganization: Combines the official FTW training and validation chips into a single 724-chip development pool, while reserving a small, independent test set exclusively for inference demonstrations. An internal validation split ($20\%$) is automatically created from the development pool during model training on the next page.

E. Verify prepared training tiles¶

Always inspect the formatted output files before starting a long training run to catch preprocessing errors early:

prepared_dir = Path(data["images_dir"]).parent

geoai.display_training_tiles(
    output_dir=str(prepared_dir),
    num_tiles=4,
    figsize=(12, 6),
    cmap="tab20",
)

During this inspection, confirm:

• Image contrast: Imagery is clear and not oversaturated or dark.

• Spatial alignment: Vector masks align perfectly with underlying imagery features.

• Instance IDs: Unique integer IDs are properly preserved across adjacent field boundaries.

• Data integrity: No corrupt tiles, missing channels, or unexpected NoData regions exist.

4. Python reactivation¶

Path("ftw_data") / "luxembourg" uses pathlib’s / operator to join path components, a cleaner alternative to os.path.join() for the same purpose. gpd.read_parquet() reads a GeoParquet file into a GeoDataFrame, the same object type you have used throughout the course for vector data, just stored in a different, more compact file format than GeoJSON or Shapefile. chips_gdf["split"].value_counts() counts how many rows fall into each unique value of the split column, a quick way to check a dataset’s class or split balance.

5. Common pitfalls¶

• Confusing the official FTW split with geoai’s internal validation split. The official train/val/test column in the metadata is combined into one training pool by prepare_ftw(); the val_split you will set during training on the next page creates a separate, fresh validation split from that pool. These are not the same split.

• Assuming imagery is already in a usable range. Raw Sentinel-2 reflectance values (0 to 10,000) are not directly usable as standard 8-bit imagery; prepare_ftw() handles this rescaling for you, but it is worth knowing it is happening.

• Treating window_a and window_b as redundant duplicates. They are deliberately different dates, included because seasonal differences carry useful information, not because of a data error.

• Skipping the visual check in step C. As in every previous data-preparation page in this course, a quick visual check catches alignment problems that are much harder to diagnose after training.

6. Mini task¶

prepare_ftw() reported “FTW luxembourg: 808 total chips, Train: 643, Val: 81, Test: 84” and “Using 724 chips for training.” Without running any code, explain where the number 724 comes from, and what happens to the 84 test chips mentioned in the first line versus the smaller number prepare_ftw() actually sets aside for testing in this tutorial.

7. Key takeaways¶

• The FTW dataset provides Sentinel-2 chips with instance-level field-boundary masks already encoded, across 24 countries, with Luxembourg used here as a compact subset.

• prepare_ftw() rescales reflectance to uint8, organizes files into the expected images/ and labels/ structure, and combines the official train and validation chips into one development pool.

• The official FTW split and geoai’s internal training-time validation split are two different things; keep them distinct.

• window_a and window_b are deliberate, different-date captures meant to help with boundary delineation through seasonal contrast.

Further reading¶

• Taylor Geospatial, Fields of The World - offcial website including a FTW detector tutorial and explorer app.

• OpenGeoAI, “Field Boundary Detection” — the full, runnable notebook this workflow is based on.

• Agricultural Field Boundary Delineation (Instance Segmentation) with GeoAI — a video walkthrough of this same dataset and task.