This repository contains information about the dataset. For information about training and models, go to: RiverScope models
Download the dataset from the following: https://zenodo.org/records/15376394 (the data is about 8GB zipped, 10GB unzipped)
Alternatively you can run the following:
wget https://zenodo.org/records/15376394/files/RiverScope.zip
📦 RiverScope
├─ PlanetScope
│ ├─ input
│ │ ├─ train (contains *.tif)
│ │ ├─ valid (contains *.tif)
│ │ └─ test (contains *.tif)
│ ├─ label
│ │ ├─ train (contains *.tif)
│ │ ├─ valid (contains *.tif)
│ │ └─ test (contains *.tif)
│ └─ derived_gt_widths-test.csv
├─ Sentinel-2
│ ├─ raw
│ │ ├─ train (contains *.tif)
│ │ ├─ valid (contains *.tif)
│ │ └─ test (contains *.tif)
│ └─ reprojected
│ ├─ train (contains *.tif)
│ ├─ valid (contains *.tif)
│ └─ test (contains *.tif)
├─ SWORD (contains *.shp)
│ └─ SWORD_nodes.csv
├─ SWOT
│ ├─ nodes (contains *.geojson)
│ └─ pixc (contains *.tif)
├─ train.csv
├─ valid.csv
└─ test.csv
- There are 4 types of data: (1) PlanetScope, (2) Sentinel-2, (3) SWORD, (4) SWOT. Each has its own folder.
- In the upper level, there are 3 csv files defining the data for each split: train/valid/test. Each csv file contains the following columns:
| Column | Data Type | Description |
|---|---|---|
| reach_id | int | The ID of the reach contained in the image |
| mid_lat | float | Latitude of the middle of the tile crop |
| mid_lon | float | Longitude of the middle of the tile crop |
| planetscope_id | str | Identifier of the PlanetScope image used for the tile. Can be used to download the raw data. |
| normalized_planetscope_path | str | Relative path of the normalized PlanetScope data, cropped to 500x500 pixels that matches the same area as the label. |
| label_path | str | Relative path of the water segmentation label for the given PlanetScope tile. [0=background, 1=river, 2=other water] |
| sword_path | str | Relative path of the SWORD vector file that contains the nodes and reaches. |
| s2_path_raw | str | Relative path of the raw Sentinel tile that contains the area captured by the PlanetScope tile |
| s2_path_reprojected | str | Relative path of the Sentinel tile that was reprojected and cropped to the same area as the PlanetScope tile |
| swot_node_path | str | Relative path of the node data contained in the given image/area |
| swot_node_id | str | Identifier of the SWOT node data, can be used to download the full,raw data |
| swot_pixc_path | str | Relative path of the rendered pixel cloud from SWOT |
| swot_pixc_id | str | Identifier of the SWOT pixel cloud, can be used to download the full,raw data |
Data type: .tif (both input and label)
PlanetScope is the high-resolution satellite image data (3m/pixel).
It has 4 bands: B,G,R,NIR (in this order) and all bands have 3m/pixel resolution.
| Band | Name |
|---|---|
| 1 | Blue |
| 2 | Green |
| 3 | Red |
| 4 | NIR |
All .tif files can be read and processed using the python library rasterio.
import rasterio
img_metadata = rasterio.open(satellite_fp) # you can use this to determine the projection, extent, etc
img_data = img_metadata.read() # this will read all available bands and will be a numpy array of shape (bands, H, W)It is a single band label with the following values per pixel:
| Value | Description |
|---|---|
| 0 | Background (non-water) |
| 1 | River water |
| 2 | Non-river water (e.g., lakes, points, etc) |
Data type: .tif
Sentinel-2 is the lower-resolution satellite image data (10m/pixel for RGB+NIR). It has 12 bands. We include two versions: the raw and the reprojected.
raw: this contains the raw, unreprojected, unnormalized Sentinel-2 bands downloaded from Sentinel Hubreprojected: this contains the sentinel images that were reprojected to be in the same view and area as the corresponding PlanetScope images. The bands are not normalized, but are interpolated from the raw bands to match the PlanetScope projection. This is useful for visually comparing Sentinel to PlanetScope. Note that the images here were bilinearly upsampled to have the same size as PlanetScope. The data in Sentinel-2 can be read in the same way as PlanetScope in the previous section.
Below are the available bands. For more information, please see here. Note that B10 is only used for atmospheric correction and is not included.
| Band | Name | Resolution (m/pixel) |
|---|---|---|
| 1 | Aerosol | 60 |
| 2 | Blue | 10 |
| 3 | Green | 10 |
| 4 | Red | 10 |
| 5 | Vegetation Red Edge | 20 |
| 6 | Vegetation Red Edge | 20 |
| 7 | Vegetation Red Edge | 20 |
| 8 | NIR | 10 |
| 8A | Narrow NIR | 20 |
| 9 | Water Vapor | 60 |
| 11 | SWIR1 | 20 |
| 12 | SWIR2 | 20 |
Data types: .shp, .cpg, .dbf, .prj, .shx
SWORD is the reference for all rivers used by SWOT and other methods. It defines nodes (points along the river spaced ~200m apart) and reaches (line segments defining rivers ~10km long). The main file is the .shp and the other extensions are supporting files. The .shp file is a vector that defines the river centerlines. If you wish you visualize the data, you can use a software QGIS. Alternatively, you can use python to read the data with the library geopandas:
import geopandas as gpd
line_gdf = gpd.read_file(sword_shp_path) # this is a geopandas dataframe that you can print
print(line_gdf) # prints all attributes as a dataframe. The 'geometry' column can be used for visualization. One line in the dataframe is one reach.
import matplotlib.pyplot as plt
fig, ax1 = plt.subplots(1, 1, figsize=(8, 8))
line_gdf.plot(ax=ax1, color='blue', label='River reach') # this visualizes the reaches contained in the shapefile.Data type: .csv
There is a csv file that contains the reference nodes available in SWORD (SWORD_nodes.csv). This supplements the raw shapefiles which contain reach-level information. This contains the node-level information such as the estimated widths.
Data types: .tif (for pixc), .geojson (for nodes)
SWOT has multiple products available. As part of our dataset we release the 3D pixel cloud (PIXC), and the node-level data (RiverSP Node). To reduce the size of our dataset, we release PIXC to be rendered in the same projection and area as the corresponding PlanetScope images. To obtain the raw files, you can refer to the swot_node_id and swot_pixc_id in the train/valid/test.csv files.
pixc: this contains the projected pixel cloud from the raw file. This is used to visualize the SWOT data before it's processed/aggregated along nodes/reaches.nodes: this contains the aggregated data frompixc. Data is aggregated into nodes (points along the river spaced around 200m apart) based on the pixels identified as water frompixc. This contains the SWOT-derived widths per node. You can read this file in python using pandas:
import pandas as pd
node_data = pd.read_csv(nodes_fp)If you found this useful, please consider citing our work:
@dataset{daroya_2025_15376394,
author = {Daroya, Rangel and
Rowley, Taylor and
Flores, Jonathan and
Friedmann, Elisa and
Bennitt, Fiona and
An, Heejin and
Simmons, Travis and
Hughes, Marissa Jean and
Kluetmeier, Camryn L and
Kica, Solomon and
Vélez, J. Daniel and
Esenther, Sarah E. and
Howard, Thomas E. and
Ye, Yanqi and
Turcotte, Audrey and
Gleason, Colin and
Maji, Subhransu},
title = {RiverScope: High-Resolution River Masking Dataset},
month = may,
year = 2025,
publisher = {Zenodo},
doi = {10.5281/zenodo.15376394},
url = {https://doi.org/10.5281/zenodo.15376394},
}