This repository provides tools that can be used to detect vessels in synthetic aperture radar (SAR) imagery produced by the Sentinel-1 satellite constellation, and electro-optical (EO) and infrared (IR) imagery produced by the Sentinel-2 satellite constellation. See: Satellite Imagery and AI: A New Era in Ocean Conservation, from Research to Deployment and Impact
These models run at global scale in production in Skylight, a free maritime intellignece platform that supports ocean conservation efforts worldwide.
The annotations are included in data/metadata.sqlite3 with the following schema. The raw data are intended to be downloaded from source (ESA), via the included download scipts.
The metadata we provide is organized according to a schema spanning a few tables. Toggle the below to read more about the details of these tables.
Schema
The sqlite database has the following tables (and associated schemas):-
collections
A collection specifies a source or catalog of imagery.
Examples:
Sentinel-1, Sentinel-2, or Landsat-8.
Schema:
cid name type notnull dflt_value pk 0 id INTEGER 0 1 1 name TEXT 0 0 Example Row:
id name 1 sentinel1 -
datasets
A dataset is a specification of a task, along with a set of labels for that task. Currently each dataset must correspond to a single collection.
Examples:
A set of point-detection labels for vessels in Sentinel-1 imagery, or a set of attribute labels for vessels within Sentinel-2 imagery, or a set of bounding-box labels for vessels in Sentinel-2 imagery.
Schema:
cid name type notnull dflt_value pk 0 id INTEGER 0 1 1 collection_id INTEGER 1 0 2 name TEXT 1 0 3 task TEXT 1 0 4 categories TEXT 0 0 Example Row:
id collection_id name task categories 1 1 vessels point ["vessel"] Notes:
Currently the only supported tasks are
point
(for point detection) andcustom
(for prediction of attributes associated with points). -
images
An image is a reference to a satellite image, optionally including additional information pertaining to how the image is intended to be processed downstream (i.e. in our repo). A single record includes information such as the name of an image file (as defined by the provider), acquisition time, file format type, location extent, and binned corner coordinates corresponding to a specified cooordinate system.
Schema:
cid name type notnull dflt_value pk 0 id INTEGER 0 1 1 uuid TEXT 1 0 2 name TEXT 1 0 3 format TEXT 1 0 4 channels TEXT 1 0 5 width INTEGER 1 0 6 height INTEGER 1 0 7 preprocessed INTEGER 1 0 0 8 hidden INTEGER 1 0 0 9 bounds TEXT 1 0 10 time DATETIME 1 0 11 projection TEXT 0 0 12 column INTEGER 1 0 13 row INTEGER 1 0 14 zoom INTEGER 1 0 Example Row:
id uuid name format channels width height preprocessed hidden bounds time projection column row zoom 1 a97d69a45df546bb8d02eb6fb794c4ef S1A_IW_GRDH_1SDV_20211104T170749_20211104T170814_040424_04CAD6_4BAE.SAFE geotiff [{"Name":"vh","Count":1},{"Name":"vv","Count":1},{"Name":"vh_overlap0","Count":1},{"Name":"vh_overlap1","Count":1},{"Name":"vh_overlap2","Count":1},{"Name":"vh_overlap3","Count":1},{"Name":"vv_overlap0","Count":1},{"Name":"vv_overlap1","Count":1},{"Name":"vv_overlap2","Count":1},{"Name":"vv_overlap3","Count":1}] 31428 24935 0 0 {"Min":{"Lon":-175.24431616814434,"Lat":-18.12758186277979},"Max":{"Lon":-172.54682929070293,"Lat":-16.082187769430707}} 2021-11-04T17:07:49+00:00 epsg:3857 55407 2287033 13 Notes:
-
There are three distinct but related coordinate systems arising from Web Mercator.
- One is of course the actual global coordinate system known as Web Mercator and referenced by epsg:3857. The unit of these coordinates is referred to as a meter; more information about maximal bounds and coverage of the coordinate system is documented here. To avoid confusion, we refer to these coordinates as absolute Web Mercator coordinates.
- Another is a coordinate system that can be used to describe discretized (or pixel) coordinates in the space of pixels on a tiling of web mercator. The process goes as follows. Shift absolute Web Mercator coordinates such that they are all positive, and the minimal x and y coordinates are 0. Next, reflect the y coordinates about their midpoint (i.e. y -> ymax - y)). Specify a _zoom level between 1 and 24, and a power of 2 called tile_size. Partition the portion of the world covered by the Web Mercator projection into (2^(zoom) x 2^(zoom) regular square tiles, with each tile being made up of (tilesize x tile_size pixels), and with the top left corner pixel of one tile coinciding with the (0, 0) coordinate of the translated and reflected absolute web mercator coordinate. We will refer to the pixel coordinates of this tiling as _pixel coordinates on a Web Mercator tiling (of tileSize 512 and zoom 13). These are the coordinates in which most of our metadata is recorded. The function described above to produce these pixel coordinates (and it's inverse) are defined explicitly in the functions
crs_to_pixel
andpixel_to_crs
of ./data/reports/dataset_report.ipynb. See here for a more comprehensive introduction to the relevant concepts. - Finally, we may sometimes have occasion to refer to the coordinates of entire tiles in the previously described tiling. Whenever we do so, we will refer to these coordinates as tile coordinates on a web mercator tiling (of zoom 13)
-
Projection indicates a coordinate system that is used to describe the image bounds, and objects relative to those bounds, in all references in our source code to the image. The only coordinate system in use here currently is Web Mercator, or epsg:3857. See here for more information about this coordinate system.
-
Zoom indicates the relative scale of a set of regular 512x512 "tiles" with which we partition the earth. A higher zoom level indicates more tiles are required to partition the earth (precisely, (2^(zoom) tiles) x (2^(zoom) tiles) are used to partition the earth). Currently, the only zoom level used throughout the metadata we provide is 13. It is relevant when considering pixel coordinates on a Web Mercator tiling.
-
UUID is a text identifier for an image. In the case of Sentinel-1, the uuids in the metadata correspond to the product uuid assigned by Copernicus API to the image. The uuids here are used by our pre-processing logic to construct filesystem directories.
-
Time corresponds to the time at which the satellite acquired the image, in ISO 8601 format.
-
Channels is used internally by the training logic, and includes both the raw channels included in the imagery product (i.e. VV, VH), as well as indications of overlap channels produced by our pre-processing logic.
-
Width, height specify the width and height of the image in pixel coordinates in the relevant Web Mercator tiling.
-
Column, row: the minimal x and y coordinates of the image extent in in pixel coordinates in the relevant Web Mercator tiling.
-
-
windows
A window is a cropped subset of an image.
Schema:
cid name type notnull dflt_value pk 0 id INTEGER 0 1 1 dataset_id INTEGER 1 0 2 image_id INTEGER 1 0 3 column INTEGER 1 0 4 row INTEGER 1 0 5 width INTEGER 1 0 6 height INTEGER 1 0 7 hidden INTEGER 1 0 0 8 split TEXT 1 'default' 0 Example Row:
id dataset_id image_id column row width height hidden split 1 1 51 2718762 1773812 945 945 0 nov-2021-point-train Notes:
- Column and row: the minimal x and y coordinates of a section of an image, in pixel coordinates in the relevant Web Mercator tiling.
- Width and height: widht and height of a section of an image, in pixel coordinates in the relevant Web Mercator tiling.
- Split: An aribtrary text label that can be applied to a window. Used to break datasets into pieces based on properties of interest (e.g. acquisition time), for e.g. train and validation sets.
-
labels
A label is metadata specifying some information about a window, such as a point that corresponds to the center of an object, or a set of numerical attributes associated with a crop.
Schema:
cid name type notnull dflt_value pk 0 id INTEGER 0 1 1 window_id INTEGER 1 0 2 column INTEGER 0 0 3 row INTEGER 0 0 4 width INTEGER 0 0 5 height INTEGER 0 0 6 extent TEXT 0 0 7 value FLOAT 0 0 8 properties TEXT 0 0 Example Row:
id window_id column row width height extent value properties 1 101 3316376 2046567 Notes:
- Column and row: the x and y coordinates of a point in a window, in pixel coordinates in the relevant Web Mercator tiling.
- Width and height: widht and height of a section of an image, in pixel coordinates in the relevant Web Mercator tiling.
- Properties: A bracket enclosed blob containing numerical properties associated with a point in an image. For example, some of the points at which vessels are labeled in this dataset have corresponding labels for attributes. These are denoted with properties set to values such as
{"Length":85,"Width":17,"Heading":71,"ShipAndCargoType":70,"Speed":0.4000000059604645}
.
If you'd like to use the Sentinel-1 dataset (e.g. to re-train our Sentinel-1 models), you'll need access to the full dataset, including both the metadata/labels provided with this repository, and the raw Sentinel-1 imagery referenced in our metadata.
Note: The raw imagery data occupies roughly 3.5TB of space.
The script download_imagery.py
allows one to acquire the raw (Sentinel-1) imagery referenced in our annotated dataset. It has
a few command line options, which you can read about in the source file, or using
python3 download_imagery.py --help
Manually
To get started in your own environment:
-
Create and activate a virtual environment with python 3.8.10,
E.g. run:
python3.8 -m venv .env source .env/bin/activate
-
Install gdal 3.4.3 on your machine.
E.g. on Ubuntu 20.04 run:
sudo apt-get update && \ sudo apt-get install software-properties-common -y && \ sudo add-apt-repository ppa:ubuntugis/ubuntugis-unstable && \ sudo apt-get update && sudo apt-get install gdal-bin libgdal-dev libpq-dev python3-gdal python3.8-venv -y
-
Install the python packages required for downloading the data via pip:
pip install -r download_requirements.txt
-
Create a Copernicus Hub account.
-
Export your user credentials and run the download script.
export COPERNICUS_USERNAME=your-username export COPERNICUS_PASSWORD=your-password python download_imagery.py \ --save_dir="./data/training_data" \ --db_path="./data/metadata.sqlite3"
Here the arg to
--save_dir
specifies the relative path to a directory in which all downloaded imagery will be stored, and--db_path
specifies the relative path to the sqlite3 metadata file distributed with this repo.
You should see all Sentinel-1 images referenced in our annotated metadata begin downloading to your specified save_dir
. The logs will indicate the status of the downloading procedure. When completed, you should have three new directories in your --save_dir
: preprocess, images, and image_cache. preprocess contains preprocessed (i.e. warped and cropped) copies of the raw Sentinel-1 imagery to which our metadata points. It is all that is strictly necessary for training. image_cache contains the raw Sentinel-1 imagery product archives.
Via Docker
You can also use our docker container to avoid setting up the necessary environment manually.
Note: Our docker images implicitly assume an x86 architechture. While you may have luck building these images on other architechtures ( e.g. by using docker's --platform
arg), we have not tested this functionality. For example, we have not tested these builds on ARM machines.
If you have trouble building the image from source on your machine, you can pull it from ghcr (assuming you have copied this repository to GitHub, and have triggered the build_and_push_docker_images
workflow from the commit of interest), try using the scripts without docker,
or use an x86 VM to build.
To use the docker image:
-
Build the docker image.
docker build . -f Dockerfile -t vessel-detection-sentinels
-
Create a Copernicus Hub account, and export your user credentials:
export COPERNICUS_USERNAME=your-username export COPERNICUS_PASSWORD=your-password
-
Run the docker image.
```bash docker run -d -v /path/to/host/dir:/home/vessel_detection/data \ -e COPERNICUS_USERNAME -e COPERNICUS_PASSWORD \ vessel-detection-download:$IMAGE_TAG \ --save_dir="/home/vessel_detection/data/training_data" \ --db_path="/home/vessel_detection/data/metadata.sqlite3" ``` As before, the arg to `--save_dir` specifies the relative path to a directory in which all downloaded imagery will be stored, and `--db_path` specifies the relative path to the sqlite3 metadata file distributed with this repo. In the example command shown here, we've pointed these to paths in the container that correspond to a mounted volume.
Once you have all Sentinel-1 imagery associated with our metadata downloaded, you can train our Sentinel-1 detection and attribute prediction models.
Note: Training currently requires GPU, due to use of torch's GradScaler
for gradient scaling. The use of
gradient scaling and automatic mixed precision can be removed from the logic easily by the interested user, but
given the affect on performance, and that the time required to train on CPU is prohibitively long in any case,
we hardcode the use of a GPU device in the training scheme.
The script train.py
allows one to train our models on our annotated dataset. It has
a few command line options, which you can read about in the source file, or using
python3 train.py --help
To train our Sentinel-1 models, we recommend using a system with:
- An x86 architecture.
- A GPU with >= 16GB memory.
- RAM >= 32GB.
These are the specs on which functionality has been tested, though they are not necessarily minimal.
To provide a sense for expected training time, we report total training time for each training run specified in the default config ./data/cfg/ on a particular machine.
Instance Type | RAM | GPU Model | GPU Memory | model type | model and dataset config | epochs | training time (hours) |
---|---|---|---|---|---|---|---|
GCP n1-highmem-8 | 52GB | NVIDIA T4 | 16GB | detector (frcnn_cmp2) | train_sentinel1_detector.json | 3 | 6 |
GCP n1-highmem-8 | 52GB | NVIDIA T4 | 16GB | attribute predictor (attr) | train_sentinel1_attribute_predictor.json | 80 | 4 |
We describe below how to set up an environment on your machine for training (either manually, or via docker).
Note: Our docker images implicitly assume an x86 architecture. While you may have luck building these images on other architechtures ( e.g. by using docker's --platform
arg),
we have not tested this functionality. For example, we have not tested these builds on ARM machines.
If you have trouble building the image from source on your machine, you can pull it from ghcr, use the scripts without docker,
or use an x86 VM to build.
Manually
To get started in your own environment:
-
Create and activate a virtual environment with python 3.8.10.
e.g. via
python3.8 -m venv .env source .env/bin/activate
-
Install gdal 3.4.3 on your system.
E.g. on Ubuntu 20.04 run:
sudo apt-get update && \ sudo apt-get install software-properties-common -y && \ sudo add-apt-repository ppa:ubuntugis/ubuntugis-unstable && \ sudo apt-get update && sudo apt-get install gdal-bin libgdal-dev libpq-dev python3-gdal python3.8-venv -y
-
Install the python packages required for training via pip:
pip install -r train_requirements.txt
-
Place the entire
preprocess
folder generated by the data download script, as well as the./data/metadata.sqlite3
file, somewhere accessible on your machine.In what follows we'll assume we've moved both of these things to the
./data
folder in this repository. -
Run the training script for the detection model.
Logging training metrics to a WANDB account:
export RUN_TAG=$(git rev-parse HEAD) export WANDB_MODE=online export WANDB_API_KEY=your-api-key python train.py \ --config_path ./data/cfg/train_sentinel1_detector.json \ --training_data_dir ./data \ --save_dir ./data/models \ --metadata_path ./data/metadata.sqlite3
Logging training metrics only locally:
export RUN_TAG=$(git rev-parse HEAD) python train.py \ --config_path ./data/cfg/train_sentinel1_detector.json \ --training_data_dir ./data \ --save_dir ./data/models \ --metadata_path ./data/metadata.sqlite3
-
Run the training script for the attribute prediction model.
Logging training metrics to a WANDB account:
export RUN_TAG=$(git rev-parse HEAD) export WANDB_MODE=online export WANDB_API_KEY=your-api-key python train.py \ --config_path ./data/cfg/train_sentinel1_attribute_predictor.json \ --training_data_dir ./data \ --save_dir ./data/models \ --metadata_path ./data/metadata.sqlite3
Logging training metrics only locally:
export RUN_TAG=$(git rev-parse HEAD) python train.py \ --config_path ./data/cfg/train_sentinel1_attribute_predictor.json \ --training_data_dir ./data \ --save_dir ./data/models \ --metadata_path ./data/metadata.sqlite3
As training progresses for each model, the logs will indicate the status of the run.
When training for a model completes, you should see the --save_dir
you specified populated with a directory named {model_name}/{RUN_TAG}
. This directory will house
saved weights, and training and model config. Here {model_name}
will indicate whether the model being trained is an instance of the detection model (the default config uses
model_name=frcnn_cmp2
for that), or an instance of the attribute prediction model (the default config uses model_name=attr
for that).
Via Docker
You can also use our docker container to avoid setting up the necessary environment manually.
-
Build the docker image.
export IMAGE_TAG=$(git rev-parse --short HEAD) docker build . -f Dockerfile -t vessel-detection-sentinels:$IMAGE_TAG
-
Place the entire
preprocess
folder generated by the data download script, as well as the./data/metadata.sqlite3
file, somewhere accessible on your machine.In what follows we'll assume we've moved both things to the
/path/to/your/data
folder on our machine. -
Run the training container for the detector:
Logging training metrics via a WANDB account:
export RUN_TAG=$(git rev-parse HEAD) export WANDB_MODE=online export WANDB_API_KEY=your-api-key docker run --shm-size 16G --gpus='"device=0"' -d \ -v /path/to/your/data:/home/vessel_detection/data \ -e RUN_TAG=$RUN_TAG \ -e WANDB_API_KEY=$WANDB_API_KEY \ -e WANDB_MODE=$WANDB_MODE \ vessel-detection-train:$IMAGE_TAG \ --config_path /home/vessel_detection/data/cfg/train_sentinel1_detector.json \ --training_data_dir /home/vessel_detection/data \ --save_dir /home/vessel_detection/data/models \ --metadata_path /home/vessel_detection/data/metadata.sqlite3
or without WANDB:
export RUN_TAG=$(git rev-parse HEAD) export WANDB_MODE=offline docker run --shm-size 16G --gpus='"device=0"' -d \ -v /path/to/your/data:/home/vessel_detection/data \ -e RUN_TAG=$RUN_TAG \ -e WANDB_MODE=$WANDB_MODE \ vessel-detection-train:$IMAGE_TAG \ --config_path /home/vessel_detection/data/cfg/train_sentinel1_detector.json \ --training_data_dir /home/vessel_detection/data \ --save_dir /home/vessel_detection/data/models \ --metadata_path /home/vessel_detection/data/metadata.sqlite3
-
Run the training container for the attribute predictor:
Logging training metrics via a WANDB account: ```bash export RUN_TAG=$(git rev-parse HEAD) export WANDB_MODE=online export WANDB_API_KEY=your-api-key docker run --shm-size 16G --gpus='"device=0"' -d \ -v /path/to/your/data:/home/vessel_detection/data \ -e RUN_TAG=$RUN_TAG \ -e WANDB_API_KEY=$WANDB_API_KEY \ -e WANDB_MODE=$WANDB_MODE \ vessel-detection-train:$IMAGE_TAG \ --config_path /home/vessel_detection/data/cfg/train_sentinel1_attribute_predictor.json \ --training_data_dir /home/vessel_detection/data \ --save_dir /home/vessel_detection/data/models \ --metadata_path /home/vessel_detection/data/metadata.sqlite3 ``` or without WANDB: ```bash export RUN_TAG=$(git rev-parse HEAD) export WANDB_MODE=offline docker run --shm-size 16G --gpus='"device=0"' -d \ -v /path/to/your/data:/home/vessel_detection/data \ -e RUN_TAG=$RUN_TAG \ -e WANDB_MODE=$WANDB_MODE \ vessel-detection-train:$IMAGE_TAG --config_path /home/vessel_detection/data/cfg/train_sentinel1_attribute_predictor.json \ --training_data_dir /home/vessel_detection/data \ --save_dir /home/vessel_detection/data/models \ --metadata_path /home/vessel_detection/data/metadata.sqlite3 ```
As training progresses for each model, the logs will indicate the status of the run.
When training for a model completes, you should see the --save_dir
you specified populated with a directory named {model_name}/{RUN_TAG}
. This directory will house
saved weights, and training and model config. Here {model_name}
will indicate whether the model being trained is an instance of the detection model (the default config uses
model_name=frcnn_cmp2
for that), or an instance of the attribute prediction model (the default config uses model_name=attr
for that).
You can perform inference right off the bat with our provided trained weights, or with weights you produce from running our training scripts.
The script src/main.py
allows one to perform inference on a raw (but de-compressed) Sentinel-1 and Sentinel-2 image products.
To perform inference with trained copies of our models, we recommend using a system with:
- An x86 architecture.
- A GPU with >= 8GB memory.
- RAM >= 32GB.
The lowest values reported here are the specs on which functionality has been tested, but are not necessarily the minimal requirements.
To provide a sense for expected inference time, we report total inference time on two particular particular machines, on representative Sentinel-1 scenes. Run time and resource requirements for our Sentinel-2 models are similar.
Instance Type | RAM | GPU Model | GPU Memory | inference time (seconds) | using cpu only |
---|---|---|---|---|---|
GCP n1-highmem-8 | 52GB | NVIDIA T4 | 16GB | 307 | false |
AMD Ryzen 9 3900X 12-Core Processor | 64GB | NVIDIA RTX 2070 SUPER | 8GB | 174 | false |
AMD Ryzen 9 3900X 12-Core Processor | 64GB | NVIDIA RTX 2070 SUPER | 8GB | 2100 | true |
We describe below how to set up an environment on your machine for inference (either manually, or via docker).
Manually
To get started in your own environment:
-
Download a Sentinel-1 scene (and optionally two historical overlaps of that scene) from the Copernicus Hub UI or API, and place it in a folder of your choosing.
For example, let's suppose we've downloaded three such scenes which we've placed on our system at:
./data/S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE
./data/S1B_IW_GRDH_1SDV_20211118T025212_20211118T025237_029636_03896D_FCBE.SAFE
./data/S1B_IW_GRDH_1SDV_20211106T025212_20211106T025237_029461_03840B_6E73.SAFE
-
Collect a directory of trained Senntinel-1 model artifacts on your machine. At minimum, you will need one set of trained weights and model config for the detection model and one set for the attribute prediction model. You can obtain these directories by training the models yourselves, or by using pre-trained weights found in data/model_artifacts.
For example, let's suppose we're using the pre-trained Sentinel-1 artifacts distributed separately, and we've copied the following directory structure with these artifacts into the `./data` folder of this repo: ``` ./data |----attr | |---c34aa37 | |-----best.pth | |-----cfg.json | |---frcnn_cmp2 | |--3dff445 |-----best.pth |-----cfg.json ```
-
Create and activate a virtual environment with python 3.8.10.
e.g. via
python3.8 -m venv .env source .env/bin/activate
-
Install gdal 3.4.3 on your system.
E.g. on Ubuntu 20.04 run:
sudo apt-get update && \ sudo apt-get install software-properties-common -y && \ sudo add-apt-repository ppa:ubuntugis/ubuntugis-unstable && \ sudo apt-get update && apt-get install gdal-bin libgdal-dev libpq-dev python3-gdal python3.8-venv -y
-
Install the python packages required for training via pip:
pip install -r requirements.txt
-
Run the inference script for the detection model.
To run inference on a single image, without access to historical overlaps:
-
Download a Sentinel-2 scene (and optionally two historical overlaps of that scene) from the Copernicus Hub UI or API, and place it in a folder of your choosing.
For example, let's suppose we've downloaded three such scenes which we've placed on our system at: - `./data/S2A_MSIL1C_20230108T060231_N0509_R091_T42RUN_20230108T062956.SAFE` - `./data/S2A_MSIL1C_20230111T061221_N0509_R134_T42RUN_20230111T064344.SAFE` - `./data/S2B_MSIL1C_20230106T061239_N0509_R134_T42RUN_20230106T063926.SAFE`
-
Collect a directory of trained Sentinel-2 model artifacts on your machine. At minimum, you will need one set of trained weights and model config for the detection model and one set for the attribute prediction model. You can obtain these directories by training the models yourselves, or by using pre-trained weights we will provide separately.
For example, let's suppose we're using the pre-trained Sentinel-2 artifacts distributed separately, and we've copied the following directory structure with these artifacts into the `./data` folder of this repo: ``` ./data |----attr | |---e609150-sentinel2-attr-resnet | |-----best.pth | |-----cfg.json | |---frcnn_cmp2 | |--15cddd5-sentinel2-swinv2-small-fpn |-----best.pth |-----cfg.json ```
-
Create and activate a virtual environment with python 3.8.10.
e.g. via
python3.8 -m venv .env source .env/bin/activate
-
Install gdal 3.4.3 on your system.
E.g. on Ubuntu 20.04 run:
sudo apt-get update && \ sudo apt-get install software-properties-common -y && \ sudo add-apt-repository ppa:ubuntugis/ubuntugis-unstable && \ sudo apt-get update && apt-get install gdal-bin libgdal-dev libpq-dev python3-gdal python3.8-venv -y
-
Install the python packages required for training via pip:
pip install -r requirements.txt
-
Run the inference script for the detection model.
To run inference on a single image, without access to historical overlaps:
Here
--raw_path
must point to a directory containing the directory specified by--scene_id
(and--historical1
and--historical2
, if they are provided).The model weight dirs for the detector and attribute predictor models are respectively pointed to by
--detector_model_dir
and--postprocess_model_dir
.The outputs of the detection script will be populated in the directory specified by
--output
.If you'd prefer to run the inference script on CPU, you can optionally pass the flag
--force-cpu
. However, please note that running inference on CPU is currently prohibitively slow. For example, while inference with a GPU might take ~2 mins for a Sentinel-1 scene, it might take closer to ~30 mins to perform the same call on CPU.
Via Docker
You can also use our docker container to avoid setting up the necessary environment manually.
Note: Our docker images implicitly assume an x86 architecture. While you may have luck building these images on other architectures ( e.g. by using docker's --platform
arg), we have not tested this functionality. For example, we have not tested these builds on ARM machines.
If you have trouble building the image from source on your machine, you can pull it from ghcr, use the scripts without docker,
or use an x86 VM to build.
-
Acquire the docker image.
Pull the latest container from Github container registry:
export REPO_OWNER=your-repo-owner-name export REPO_NAME=sentinel-vessel-detection export IMAGE_TAG=latest docker pull ghcr.io/$REPO_OWNER/$REPO_NAME/vessel-detection:$IMAGE_TAG
or build it from source by running the following from the top level of this repo:
export IMAGE_TAG=$(git rev-parse --short HEAD) docker build . -f Dockerfile -t vessel-detection:$IMAGE_TAG
-
Prepare a machine with at least 16GB RAM, and a GPU w/ >= 8GB memory.
-
Download a Sentinel-1 scene (and optionally up to two historical overlaps) from the Copernicus Hub UI or API, and place it in a folder of your choosing.
For example, let's suppose we've downloaded three such scenes which we've placed on our system at:
/path/to/your/data/S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE
/path/to/your/data/S1B_IW_GRDH_1SDV_20211118T025212_20211118T025237_029636_03896D_FCBE.SAFE
/path/to/your/data/S1B_IW_GRDH_1SDV_20211106T025212_20211106T025237_029461_03840B_6E73.SAFE
-
Collect a directory of trained Sentinel-1 model artifacts on your machine. At minimum, you will need one set of trained weights and model config for the detection model and one set for the attribute prediction model. You can obtain these directories by training the models yourselves, or by using pre-trained weights we will provide separately.
For example, let's suppose we're using the pre-trained artifacts distributed separately, and we've copied the following directory structure with these artifacts into the `/path/to/your/data` folder of this repo: ``` /path/to/your/data |----attr | |---c34aa37 | |-----best.pth | |-----cfg.json | |---frcnn_cmp2 | |--3dff445 |-----best.pth |-----cfg.json ```
-
On that machine,
To run inference on a single image, without access to historical overlaps, run:
docker run --shm-size 16G --gpus='"device=0"' \ -v /path/to/your/data:/home/vessel_detection/data vessel-detection:$IMAGE_TAG \ --raw_path=/home/vessel_detection/data/ \ --scratch_path=/home/vessel_detection/data/scratch/ \ --output=/home/vessel_detection/data/output/ \ --detector_model_dir=/home/vessel_detection/data/models/frcnn_cmp2/3dff445 \ --postprocess_model_dir=/home/vessel_detection/data/models/attr/c34aa37 \ --scene_id=S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE \ --conf=.9 --nms_thresh=10 --save_crops=True --catalog=sentinel1
To run inference on a single image, using additional context from a single historical overlap of the image:
docker run --shm-size 16G --gpus='"device=0"' \ -v /path/to/your/data:/home/vessel_detection/data vessel-detection:$IMAGE_TAG \ --raw_path=/home/vessel_detection/data/ \ --scratch_path=/home/vessel_detection/data/scratch/ \ --output=/home/vessel_detection/data/output/ \ --detector_model_dir=/home/vessel_detection/data/models/frcnn_cmp2/3dff445 \ --postprocess_model_dir=/home/vessel_detection/data/models/attr/c34aa37 \ --historical1=S1B_IW_GRDH_1SDV_20211118T025212_20211118T025237_029636_03896D_FCBE.SAFE \ --scene_id=S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE \ --conf=.9 --nms_thresh=10 --save_crops=True --catalog=sentinel1
To run inference on a single image, using additional context from two historical overlaps of the image:
docker run --shm-size 16G --gpus='"device=0"' \ -v /path/to/your/data:/home/vessel_detection/data vessel-detection:$IMAGE_TAG \ --raw_path=/home/vessel_detection/data/ \ --scratch_path=/home/vessel_detection/data/scratch/ \ --output=/home/vessel_detection/data/output/ \ --detector_model_dir=/home/vessel_detection/data/models/frcnn_cmp2/3dff445 \ --postprocess_model_dir=/home/vessel_detection/data/models/attr/c34aa37 \ --historical1=S1B_IW_GRDH_1SDV_20211118T025212_20211118T025237_029636_03896D_FCBE.SAFE \ --historical2=S1B_IW_GRDH_1SDV_20211106T025212_20211106T025237_029461_03840B_6E73.SAFE \ --scene_id=S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE \ --conf=.9 --nms_thresh=10 --save_crops=True --catalog=sentinel1
-
Acquire the docker image.
Pull the latest container from Github container registry:
export REPO_OWNER=your-repo-owner-name export REPO_NAME=sentinel-vessel-detection export IMAGE_TAG=latest docker pull ghcr.io/$REPO_OWNER/$REPO_NAME/vessel-detection:$IMAGE_TAG
or build it from source by running the following from the top level of this repo:
export IMAGE_TAG=$(git rev-parse --short HEAD) docker build . -f docker/inference.dockerfile -t vessel-detection:$IMAGE_TAG
-
Prepare a machine with at least 16GB RAM, and a GPU w/ >= 8GB memory.
-
Download a Sentinel-2 scene (and optionally up to two historical overlaps) from the Copernicus Hub UI or API, and place it in a folder of your choosing.
For example, let's suppose we've downloaded three such scenes which we've placed on our system at:
/path/to/your/data/S2A_MSIL1C_20230108T060231_N0509_R091_T42RUN_20230108T062956.SAFE
/path/to/your/data/S2A_MSIL1C_20230111T061221_N0509_R134_T42RUN_20230111T064344.SAFE
/path/to/your/data/S2B_MSIL1C_20230106T061239_N0509_R134_T42RUN_20230106T063926.SAFE
-
Collect a directory of trained Sentinel-2 model artifacts on your machine. At minimum, you will need one set of trained weights and model config for the detection model and one set for the attribute prediction model. You can obtain these directories by training the models yourselves, or by using pre-trained weights we will provide separately.
For example, let's suppose we're using the pre-trained artifacts distributed separately, and we've copied the following directory structure with these artifacts to `path/to/your/data` folder on your machine: ``` path/to/your/data |----attr | |---e609150-sentinel2-attr-resnet | |-----best.pth | |-----cfg.json | |---frcnn_cmp2 | |--15cddd5-sentinel2-swinv2-small-fpn |-----best.pth |-----cfg.json ```
-
On that machine,
To run inference on a single image, without access to historical overlaps, run: ```bash docker run --shm-size 16G --gpus='"device=0"' \ -v /path/to/your/data:/home/vessel_detection/data vessel-detection:$IMAGE_TAG \ --raw_path=/home/vessel_detection/data/ \ --scratch_path=/home/vessel_detection/data/scratch/ \ --output=/home/vessel_detection/data/output/ \ --detector_model_dir=/home/vessel_detection/data/models/frcnn_cmp2/15cddd5-sentinel2-swinv2-small-fpn \ --postprocess_model_dir=/home/vessel_detection/data/models/attr/e609150-sentinel2-attr-resnet \ --scene_id=S2A_MSIL1C_20230108T060231_N0509_R091_T42RUN_20230108T062956.SAFE \ --conf=.9 --nms_thresh=10 --save_crops=True --catalog=sentinel2 ``` To run inference on a single image, using additional context from a single historical overlap of the image: ```bash docker run --shm-size 16G --gpus='"device=0"' \ -v /path/to/your/data:/home/vessel_detection/data vessel-detection:$IMAGE_TAG \ --raw_path=/home/vessel_detection/data/ \ --scratch_path=/home/vessel_detection/data/scratch/ \ --output=/home/vessel_detection/data/output/ \ --detector_model_dir=/home/vessel_detection/data/models/frcnn_cmp2/15cddd5-sentinel2-swinv2-small-fpn \ --postprocess_model_dir=/home/vessel_detection/data/models/attr/e609150-sentinel2-attr-resnet \ --historical1=S2A_MSIL1C_20230111T061221_N0509_R134_T42RUN_20230111T064344.SAFE \ --scene_id=S2A_MSIL1C_20230108T060231_N0509_R091_T42RUN_20230108T062956.SAFE \ --conf=.9 --nms_thresh=10 --save_crops=True --catalog=sentinel2 ``` To run inference on a single image, using additional context from two historical overlaps of the image: ```bash docker run --shm-size 16G --gpus='"device=0"' \ -v /path/to/your/data:/home/vessel_detection/data vessel-detection:$IMAGE_TAG \ --raw_path=/home/vessel_detection/data/ \ --scratch_path=/home/vessel_detection/data/scratch/ \ --output=/home/vessel_detection/data/output/ \ --detector_model_dir=/home/vessel_detection/data/models/frcnn_cmp2/15cddd5-sentinel2-swinv2-small-fpn \ --postprocess_model_dir=/home/vessel_detection/data/models/attr/e609150-sentinel2-attr-resnet \ --historical1=S2A_MSIL1C_20230111T061221_N0509_R134_T42RUN_20230111T064344.SAFE \ --historical2=S2B_MSIL1C_20230106T061239_N0509_R134_T42RUN_20230106T063926.SAFE \ --scene_id=S2A_MSIL1C_20230108T060231_N0509_R091_T42RUN_20230108T062956.SAFE \ --conf=.9 --nms_thresh=10 --save_crops=True --catalog=sentinel2 ``` ### Notes (common to both imagery sources): Here `--raw_path` must point to a directory containing the file specified by `--scene_id` (and ``--historical1`` and ``--historical2``, if they are provided). The model weight dirs for the detector and attribute predictor models are respectively pointed to by `--detector_model_dir` and `--postprocess_model_dir`. The outputs of the detection script will be populated in the directory specified by `--output`. If you'd prefer to run the inference script on CPU, you can optionally pass the flag `--force-cpu`. However, please note that running inference on CPU is currently prohibitively slow. For example, while inference with a GPU might take ~2 mins for a Sentinel-1 scene, it might take closer to ~30 mins to perform the same call on CPU.
Running the inference script will produce at most three types of artifact in the specified output directories:
-
A csv,
predictions.csv
, with one row for each detected vessel.Here is the schema of the output csv:
column name description preprocess_row The x coordinate of the detection, in pixel coordinates in the preprocessed imagery (saved to the output scratch dir). preprocess_column The y coordinate of the detection, in pixel coordinates in the preprocessed imagery (saved to the output scratch dir). lat Latitude associated with the detection. lon Longitude associated with the detection. score Float between 0 and 1 representing the confidence that a vessel was detected at the specified location in the image. vessel_length_m Vessel length (meters). vessel_width_m Vessel width (meters). heading_bucket_i Probability that the heading direction of the detected vessel lies in the range between i*(360/16) and (i + 1)* (360/16) degrees, measured clockwise relative to true north. vessel_speed_k Vessel speed (knots). is_fishing_vessel Float between 0 and 1 representing the probability that the detected vessel is a fishing vessel. detect_id Same as scene*id below, but with a zero-indexed *{idx}
string appended to the name, specifying the count of detections at the time this row was written.scene_id The provider defined name of the scene on which detection was performed. See e.g. https://sentinels.copernicus.eu/web/sentinel/technical-guides/sentinel-1-sar/products-algorithms/level-1-product-formatting for Sentinel-1 and https://sentinels.copernicus.eu/web/sentinel/user-guides/sentinel-2-msi/naming-convention for Sentinel-2. row The x coordinate of the detection, in pixel coordinates in the original imagery. column The y coordinate of the detection, in pixel coordinates in the original imagery . orientation The clockwise rotation angle (in degrees) necessary to rotate an upwards pointing vertical line in the image to align with North meters_per_pixel The spatial resolution (in meters) of a single pixel in the output crops associated with a detection. -
Cropped PNG images (of size 128x128) surrounding each vessel detection, for certain subsets of spectral channels in the input imagery. These are only produced if inference is run with
--save_crops
set to True. If produced, these files are named according to the convention:{scene_id}_{detection_idx}_{channel_subset_abbreviation}.png
.- For Sentinel-1, we output two crops, one for each of the imagery channels. Here is a sample of the VH and VV polarization channels for the cropped outputs of a vessel detected in Sentinel-1 imagery: <figure><img src=./data/example_output/sentinel1/S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE_0_vh.png alt="Detection Crop in VH Channel" style="width:48%"> <img src=./data/example_output/sentinel1/S1B_IW_GRDH_1SDV_20211130T025211_20211130T025236_029811_038EEF_D350.SAFE_0_vv.png alt="Detection Crop in VV Channel" style="width:48%"><br></figure> <figcaption align = "center"><b>A Sentinel-1 vessel detection</b>. <br> <b>Left: VH polarization. Right: VV polarization.</b></figcaption> <br> - For Sentinel-2, we output at most 4 crops, one for the TCI channel collection, and one each for bands 8, 11, 12, if they are used by the model in question. Here is a sample of the TCI image corresponding to the cropped output of a vessel detected in Sentinel-2 imagery: <figure align = "center" ><img src=./data/example_output/sentinel2/S2B_MSIL1C_20211102T061019_N0301_R134_T42RUN_20211102T071305.SAFE_0_tci.png alt="Detection Crop in TCI channel" style="width:60%"> <figcaption align = "center"><b>A Sentinel-2 vessel detection</b>. <br> <b> TCI Channels.</b></figcaption> <br> </figure>
-
Warped copies of the relevant TIFF files in the input, and numpy arrays containing processed variants of the input images, in the scratch folder specified.
By default, these intermediate files get produced in the specified scratch directory at runtime, and then cleaned up at the end of the inference call. If you'd like to keep these around to inspect them or use them in some way, pass the flag `--keep_intermediate_files` to the detect call.
- Skylight-ML (especially Mike Gartner who wrote most of this codebase
- PRIOR at AI2 (especially Favyen Bastani and Piper Wolters) contributed considerable expertise to the foundational architectures of both models
- European Space Agency for making Sentinel-1 and Sentinel-2 data available to the public.
- The Defense Innovation Unit (DIU) who built the foundation for this work through xView3.