Maldives Ecosystem Mapping

As part of the Global Ecosystem Atlas Maldives Accelerator project (https://earthobservations.org/solutions/incubators/global-ecosystems-atlas), Ai2 has worked with partners to create the first version of an ecosystem category segmentation model that can be used to map terrestrial and near-land ecosystems in the Maldives.

This repository contains the workflow and configuration files needed to reproduce the training of the ecosystem category segmentation model, and supports applying the model on new images.

We have also released the ecosystem map of the Maldives computed by the model; see the Data Access Information section below for how to download this data.

Training

Requirements

• Python (3.12 is recommended)
• GPU

Note that testing was done on a Google Compute Engine VM with the following specifications:

• n1-standard-8 (8 vCPUs and 30 GB RAM)
• GPU: 1 x NVIDIA T4
• Image: Deep Learning VM with CUDA 11.8, M125 (Debian 11, Python 3.10, CUDA 11.8)

Download Dataset

To train and apply the model, first obtain the rslearn dataset from the release bucket on Google Cloud Storage:

export DATASET_PATH=/local/dataset/path/
mkdir $DATASET_PATH
cd $DATASET_PATH
wget https://storage.googleapis.com/ai2-earthsystem-release/Release/GlobalEcosystemSegmenter/202410/rslearn_dataset.zip
unzip rslearn_dataset.zip

Above, replace /local/dataset/path/ with the local path of your choice.

The dataset contains:

• Sentinel-2 images spanning each island in the Maldives. There are six images per island, with all being used by the model, to provide robustness to cloud cover and other artifacts.
• Crops of the images that were annotated with segmentation labels.
• The corresponding labels.

The image crops and labels are used to train and validate the model, while the island-spanning images are used afterward to obtain predictions over the entire Maldives.

The dataset is intended for use with rslearn, AI2's remote sensing model development tool. Training and prediction examples are stored in the windows/ directory; each example is a "spatiotemporal window" with a defined projection, spatial bounds, and time range, and includes "layers" containing images and labels. They are divided into four groups of windows:

• images_sentinel2: Sentinel-2 images of entire islands.
• crops_sentinel2: the portions of islands that have segmentation labels.
• images_planetscope: windows that reference commercial PlanetScope images of entire islands. The windows are included so that you can easily download the commercial images using rslearn in case you have a Planet Labs account (requires purchase).
• crops_planetscope: portions of islands with segmentation labels, coupled with PlanetScope image crops.

The labels (windows/crops_sentinel2/*/layers/label/label/geotiff.tif) are GeoTIFFs with integer values specifying an ecosystem class:

0. Unknown
1. FM_1_3_INTERMITTENTLY_CLOSED_AND_OPEN_LAKES_AND_LAGOONS
2. F_2_2_SMALL_PERMANENT_FRESHWATER_LAKES
3. MFT_1_2_INTERTIDAL_FORESTS_AND_SHRUBLANDS
4. MFT_1_3_COASTAL_SALTMARSHES_AND_REEDBEDS
5. MT_1_1_ROCKY_SHORELINES
6. MT_1_3_SANDY_SHORELINES
7. MT_2_1_COASTAL_SHRUBLANDS_AND_GRASSLANDS
8. MT_3_1_ARTIFICIAL_SHORELINES
9. M_1_1_SEAGRASS_MEADOWS
10. M_1_3_PHOTIC_CORAL_REEFS
11. M_1_6_SUBTIDAL_ROCKY_REEFS
12. M_1_7_SUBTIDAL_SAND_BEDS
13. TF_1_3_PERMANENT_MARSHES
14. T_7_1_ANNUAL_CROPLANDS
15. T_7_3_PLANTATIONS
16. T_7_4_URBAN_AND_INDUSTRIAL_ECOSYSTEMS

You can open Sentinel-2 images and corresponding labels in GIS software like qgis, e.g.:

qgis $DATASET_PATH/windows/crops_sentinel2/HaaAlifu_Baarah_LD0987_30081_-75425_sentinel2/layers/sentinel2/B02_B03_B04_B08/geotiff.tif $DATASET_PATH/windows/crops_sentinel2/HaaAlifu_Baarah_LD0987_30081_-75425_sentinel2/layers/label/label/geotiff.tif

Install rslearn

git clone https://github.com/allenai/rslearn.git
python3 -m venv your_env_name
source your_env_name/bin/activate
cd rslearn
pip install .[extra]

Train the model:

cd /path/to/maldives_ecosystem_mapping/
rslearn model fit --config model_configs/config_sentinel2.yaml --data.init_args.path $DATASET_PATH

After training, find the best checkpoint and copy it for convenience:

cp lightning_logs/version_0/checkpoints/epoch*ckpt best.ckpt

You may need to change the version_0 to the correct subfolder in lightning_logs if you ran model fit multiple times.

As an alternative to training the model, you can download the released Sentinel-2 model checkpoint:

wget https://storage.googleapis.com/ai2-earthsystem-release/Release/GlobalEcosystemSegmenter/202410/weights/sentinel2.ckpt -O best.ckpt

You can visualize the model's outputs over the validation set:

mkdir vis
rslearn model test --config model_configs/config_sentinel2.yaml --data.init_args.path $DATASET_PATH --model.init_args.visualize_dir vis/ --ckpt_path best.ckpt

Prediction

Now we can generate GeoTIFF outputs across all the islands:

rslearn model predict --config model_configs/config_sentinel2.yaml --data.init_args.path $DATASET_PATH --ckpt_path best.ckpt --data.init_args.num_workers 8

The outputs will appear in paths like windows/images_sentinel2/*/layers/output/output/geotiff.tif (overwriting the outputs included in the rslearn dataset download).

qgis $DATASET_PATH/windows/images_sentinel2/HaaAlifu_Baarah_LD0987_sentinel2/layers/sentinel2/B02_B03_B04_B08/geotiff.tif $DATASET_PATH/windows/images_sentinel2/HaaAlifu_Baarah_LD0987_sentinel2/layers/output/output/geotiff.tif

Obtaining Images

The dataset includes Sentinel-2 images, but if you want to obtain PlanetScope images, or get new Sentinel-2 images, first swap out the dataset config with one that specifies a data source (which is used by rslearn to automatically retrieve images).

cp dataset_configs/config_planetscope.json $DATASET_PATH/config.json
cp dataset_configs/config_sentinel2.json $DATASET_PATH/config.json

Then use rslearn to populate the dataset:

rslearn dataset prepare --root $DATASET_PATH --workers 8 --group crops_planetscope
rslearn dataset ingest --root $DATASET_PATH --workers 8 --group crops_planetscope
rslearn dataset materialize --root $DATASET_PATH --workers 8 --group crops_planetscope

For PlanetScope, the PL_API_KEY environment variable must be set.

Before training the model, swap the configuration back:

cp dataset_configs/config_train.json $DATASET_PATH/config.json

Apply the Model on a New Location

The model is only trained on ecosystem labels in the Maldives, so it likely will not give accurate results in other regions, but it is nevertheless possible to apply the model elsewhere.

Start by adding a new window to rslearn. Here we put the window in a different group called "new_locations".

rslearn dataset add_windows --root $DATASET_PATH --group new_locations --utm --resolution 10 --src_crs EPSG:4326 --box=-122.414,47.587,-122.286,47.704 --start 2024-05-01T00:00:00+00:00 --end 2024-10-01T00:00:00+00:00 --name name

Replace the box with coordinates corresponding to the desired location. Then obtain images here:

cp dataset_configs/config_sentinel2.json $DATASET_PATH/config.json
rslearn dataset prepare --root $DATASET_PATH --workers 8 --group new_locations
rslearn dataset ingest --root $DATASET_PATH --workers 8 --group new_locations --no-use-initial-job --jobs-per-process 1
rslearn dataset materialize --root $DATASET_PATH --workers 8 --group new_locations --no-use-initial-job

Then apply the model:

cp dataset_configs/config_train.json $DATASET_PATH/config.json
rslearn model predict --config model_configs/config_sentinel2.yaml --data.init_args.path $DATASET_PATH --ckpt_path best.ckpt --data.init_args.predict_config.groups '["new_locations"]'

Visualize in qgis:

qgis $DATASET_PATH/windows/new_locations/name/layers/sentinel2/B02_B03_B04_B08/geotiff.tif $DATASET_PATH/windows/new_locations/name/layers/output/output/geotiff.tif

Data Access Information

All released data is located in a public Google Cloud Storage (GCS) bucket named ai2-earthsystem-release. See https://cloud.google.com/storage/docs/downloading-objects#cli-download-object to learn more about how to access GCS.

Besides the rslearn dataset (which contains training data as well as Sentinel-2 images of each island) and Sentinel-2 model checkpoint mentioned above, the following data is also available.

Vector Annotations

While the rslearn dataset contains the annotations rasterized as GeoTIFF files, the original vector annotations are also available at gs://ai2-earthsystem-release/Release/GlobalEcosystemSegmenter/202410/ in GeoJSON format.

The files are named [IslandName]_[DateOfCollect]_labels.geojson. For example, gs://ai2-earthsystem-release/Release/GlobalEcosystemSegmenter/202410/TrainingData/baarah_2024-02-24-05-48_labels.geojson contains annotations for the island named Baarah, and a portion of this particular island was labeled using an image that was collected on 2024-02-24.

Predicted Ecosystem Maps

The ecosystem maps predicted by the models in the Maldives are available at gs://ai2-earthsystem-release/Release/GlobalEcosytemAtlas/latest/Maldives/.

In the future, older versions of the data will be named by release date, e.g. the current release is gs://ai2-earthsystem-release/Release/GlobalEcosytemAtlas/20241010/.

The maps are stored as single-band GeoTIFFs, similar to the ecosystem class label GeoTIFFs described under Download Dataset above.

For each island, we compute outputs from two models, each trained on satellite images from a different sensor:

• Sentinel-2: We have processed 1,247 islands and placed data in the folder named sentinel2 (e.g., gs://ai2-earthsystem-release/Release/GlobalEcosytemAtlas/latest/Maldives/sentinel2/Baarah.tif). Note that the number of islands processed is less than PlanetScope because some islands are smaller than the resolution of some Sentinel-2 bands (i.e., less than 40 m x 40 m).
• PlanetScope: We have processed 1,419 islands and placed data in the folder named planetscope (e.g., gs://ai2-earthsystem-release/Release/GlobalEcosytemAtlas/latest/Maldives/planetscope/Baarah.tif).

The files are named based on island names provided by the Maldives in this GeoJSON: gs://ai2-earthsystem-release/Release/GlobalEcosytemAtlas/202410/Maldives/island_geo.geojson. For unnamed islands, the files are named based on the FCODE property in the GeoJSON instead.

Model Checkpoints

A PlanetScope model checkpoint is available in addition to the Sentinel-2 checkpoint:

• Sentinel-2: gs://ai2-earthsystem-release/Release/GlobalEcosystemSegmenter/202410/weights/sentinel2.ckpt
• PlanetScope: gs://ai2-earthsystem-release/Release/GlobalEcosystemSegmenter/202410/weights/planetscope.ckpt

License

• Code, model configuration files, and model checkpoints: Apache License 2.0
• Annotations: CC0 1.0 (https://creativecommons.org/publicdomain/zero/1.0/)
• Ecosystem map predictions: Open Data Commons Attribution License v1.0 (https://opendatacommons.org/licenses/by/1-0/)

Contact

For questions and suggestions, please open an issue on GitHub.